Object-oriented Analysis And Design: Understanding System Development With UML 2.0

Object-Oriented Analysis and Design Understanding System Development with UML 2.0 Mike O’Docherty Object-Oriented An...

Author: Mike O'Docherty

117 downloads 924 Views 8MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

Object-Oriented Analysis and Design Understanding System Development with UML 2.0

Mike O’Docherty

Object-Oriented Analysis and Design

Object-Oriented Analysis and Design Understanding System Development with UML 2.0

Mike O’Docherty

Copyright  2005

John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex PO19 8SQ, England Telephone (+44) 1243 779777

Email (for orders and customer service enquiries): [email protected] Visit our Home Page on www.wiley.com All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning or otherwise, except under the terms of the Copyright, Designs and Patents Act 1988 or under the terms of a licence issued by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1T 4LP, UK, without the permission in writing of the Publisher. Requests to the Publisher should be addressed to the Permissions Department, John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex PO19 8SQ, England, or emailed to [email protected], or faxed to (+44) 1243 770620. This publication is designed to provide accurate and authoritative information in regard to the subject matter covered. It is sold on the understanding that the Publisher is not engaged in rendering professional services. If professional advice or other expert assistance is required, the services of a competent professional should be sought. Other Wiley Editorial Ofﬁces John Wiley & Sons Inc., 111 River Street, Hoboken, NJ 07030, USA Jossey-Bass, 989 Market Street, San Francisco, CA 94103-1741, USA Wiley-VCH Verlag GmbH, Boschstr. 12, D-69469 Weinheim, Germany John Wiley & Sons Australia Ltd, 33 Park Road, Milton, Queensland 4064, Australia John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-01, Jin Xing Distripark, Singapore 129809 John Wiley & Sons Canada Ltd, 22 Worcester Road, Etobicoke, Ontario, Canada M9W 1L1 Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic books. Library of Congress Cataloging-in-Publication Data O’Docherty, Mike. Object-oriented analysis and design : understanding system development with UML 2.0 / Mike O’Docherty. p. cm. Includes bibliographical references and index. ISBN-13 978-0-470-09240-8 ISBN-10 0-470-09240-8 (pbk. : alk. paper) 1. Object-oriented programming (Computer science) 2. Computer software–Development. I. Title. QA76.64.O35 2005 005.1 17 – dc22 2005004182 British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library ISBN-13 978-0-470-09240-8 ISBN-10 0-470-09240-8 Typeset in 10/14 Berkeley-Medium by Laserwords Private Limited, Chennai, India Printed and bound in Great Britain by Biddles Ltd, King’s Lynn This book is printed on acid-free paper responsibly manufactured from sustainable forestry in which at least two trees are planted for each one used for paper production.

For Alice and George

Contents

1 Introduction 1.1 Background 1.2 A Brief History of Programming 1.3 Methodologies 1.4 About this Book 1.4.1 Content Overview 1.4.2 Case Studies 1.4.3 Navigation

2 4 4 5 5 6 7 7

Part I Setting the Scene

9

2 Object Concepts 2.1 Introduction 2.2 What Is an Object? 2.3 Identical or Equal? 2.4 Depicting Objects 2.5 Encapsulation 2.6 Association and Aggregation 2.7 Graphs and Trees 2.8 Links and Navigability 2.9 Messages 2.10 Invoking an Operation 2.11 An Example Collaboration 2.12 How an Object-Oriented Program Works 2.13 Garbage Collection 2.14 Classes 2.15 What Does a Class Deﬁne? 2.16 Shared Data and Shared Operations 2.17 Types 2.18 Terminology 2.19 Reusing Code

10 12 13 16 19 20 21 23 25 27 28 30 32 32 34 37 39 40 41 43

viii

Contents

2.20 Summary Further Reading Review Questions Answers to Activity 1 Answers to Review Questions

48 48 48 50 50

3 Inheritance 3.1 Introduction 3.2 Designing a Class Hierarchy 3.3 Adding Implementations to a Class Hierarchy 3.4 Abstract Classes 3.5 Redeﬁning Methods 3.6 Implementing a Stack Class 3.6.1 Implementing a Stack using Inheritance 3.6.2 Implementing a Stack using Composition 3.6.3 Inheritance versus Composition 3.7 Multiple Inheritance 3.8 Guidelines for Using Inheritance 3.9 Summary Further Reading Review Questions Answers to Review Questions

52 54 55 58 59 63 64 65 66 68 69 73 73 74 74 75

4 Type Systems 4.1 Introduction 4.2 Dynamic and Static Type Systems 4.3 Polymorphism 4.3.1 Polymorphic Variables 4.3.2 Polymorphic Messages 4.4 Dynamic Binding 4.5 Polymorphism Guideline 4.6 Type Casting 4.7 Explicit Casts 4.8 Genericity with Templates 4.9 Summary Further Reading Review Questions Answers to Activity 2 Answers to Activity 3 Answers to Review Questions

78 80 80 82 82 83 85 87 88 89 91 93 93 93 95 95 96

Contents

5 Software Development Methodologies 5.1 Introduction 5.2 Classical Phases in Software Production 5.2.1 Requirements 5.2.2 Analysis 5.2.3 Design 5.2.4 Speciﬁcation 5.2.5 Implementation 5.2.6 Testing 5.2.7 Deployment 5.2.8 Maintenance 5.2.9 Key Questions 5.3 Software Engineering and the Waterfall Methodology 5.4 New Methodologies 5.4.1 Spiral Methodology 5.4.2 Iterative Methodology 5.4.3 Incremental Methodology 5.4.4 Combining the Methodologies 5.5 Object-Oriented Methodologies 5.5.1 UML, RUP and XP 5.5.2 The Need for Development Tools 5.6 Ripple Overview 5.6.1 Use Case Diagram 5.6.2 Class Diagram (Analysis Level) 5.6.3 Communication Diagram 5.6.4 Deployment Diagram 5.6.5 Class Diagram (Design Level) 5.6.6 Sequence Diagram 5.7 Summary Further Reading Review Questions Answers to Review Questions

Part II Understanding the Problem 6 Gathering Requirements 6.1 Introduction 6.2 The Birth of a System

98 100 102 102 102 103 103 104 104 104 104 105 106 110 110 111 112 113 114 115 116 117 120 121 122 123 124 125 126 126 127 128

129 130 132 133

ix

x

Contents

6.3 6.4

Use Cases Business Perspective 6.4.1 Identifying Business Actors 6.4.2 Writing the Project Glossary 6.4.3 Identifying Business Use Cases 6.4.4 Illustrating Use Cases on a Communication Diagram 6.4.5 Illustrating Use Cases on an Activity Diagram 6.5 Developer Perspective 6.5.1 Specializing Actors 6.5.2 Use Case Relationships 6.5.3 System Use Case Details 6.5.4 Preconditions, Postconditions and Inheritance 6.5.5 Supplementary Requirements 6.5.6 User Interface Sketches 6.5.7 Prioritizing System Use Cases 6.6 Summary Further Reading Review Questions Answers to Review Questions 7 Analyzing the Problem 7.1 Introduction 7.2 Why Do Analysis? 7.3 Overview of the Analysis Process 7.4 Static Analysis 7.4.1 Finding Classes 7.4.2 Identifying Class Relationships 7.4.3 Drawing Class and Object Diagrams 7.4.4 Drawing Relationships 7.4.5 Attributes 7.4.6 Association Classes 7.4.7 Tangible versus Intangible Objects 7.4.8 Good Objects 7.5 Dynamic Analysis 7.5.1 Drawing Use Case Realizations 7.5.2 Boundaries, Controllers and Entities 7.5.3 Communication Diagram Elements 7.5.4 Adding Operations to Classes 7.5.5 Responsibilities

135 136 137 138 139 141 143 145 149 150 155 156 158 158 159 162 162 163 165 166 168 168 169 170 170 171 172 173 178 182 183 188 188 189 191 192 194 194

Contents

7.5.6 State Modeling 7.6 Summary Further Reading Review Questions Answers to Activity 4 Answers to Review Questions

Part III Designing the Solution

195 197 197 197 201 201

203

8 Designing the System Architecture 8.1 Introduction 8.2 Design Priorities 8.3 Steps in System Design 8.4 Choosing a Networked System Topology 8.4.1 The History of Network Architectures 8.4.2 Three-Tier Architecture 8.4.3 Personal Computers 8.4.4 Network Computers 8.4.5 The Internet and the World Wide Web 8.4.6 Intranets 8.4.7 Extranets and Virtual Private Networks 8.4.8 Client–Server versus Distributed Architectures 8.4.9 Depicting Network Topology in UML 8.5 Designing for Concurrency 8.6 Designing for Security 8.6.1 Digital Encryption and Decryption 8.6.2 General Security Rules 8.7 Partitioning Software 8.7.1 Systems and Subsystems 8.7.2 Layers 8.7.3 Java Layers: Applet plus RMI 8.7.4 Message Flow in Layers 8.8 Summary Further Reading Review Questions Answers to Review Questions

204 206 207 207 208 208 210 212 213 214 215 215 216 218 220 222 223 225 225 226 227 230 233 237 238 238 239

9 Choosing Technologies 9.1 Introduction

240 242

xi

xii

Contents

9.2 9.3 9.4 9.5 9.6 9.7

Client Tier Technologies Client Tier to Middle Tier Protocols Middle Tier Technologies Middle Tier to Data Tier Technologies Other Technologies Typical Front-End Conﬁgurations 9.7.1 HTML/CGI-with-Scripts 9.7.2 HTML/CGI-with-Servlets 9.7.3 RMI 9.7.4 CORBA 9.7.5 EJB 9.8 Back-End Conﬁgurations 9.9 Java E-Commerce Conﬁguration 9.10 UML Packages 9.11 Summary Further Reading Review Questions Answers to Review Questions 10 Designing the Subsystems 10.1 Introduction 10.2 Mapping the Analysis Class Model into the Design Class Model 10.2.1 Mapping Operations 10.2.2 Variable Types 10.2.3 Visibility of Fields 10.2.4 Accessors 10.2.5 Mapping Classes, Attributes and Compositions 10.2.6 Mapping Other Types of Association 10.2.7 Universal Identiﬁer 10.3 Handling Persistence with a Relational Database 10.3.1 Database Management Systems 10.3.2 The Relational Model 10.3.3 Mapping Entity Classes 10.3.4 Mapping Associations 10.3.5 Mapping Object State 10.4 Finalizing the User Interfaces 10.5 Designing the Business Services 10.5.1 Using Proxies and Copies 10.5.2 Classifying Business Services

242 244 246 247 248 250 250 252 254 255 256 258 258 262 267 267 267 268 270 272 273 273 274 274 275 276 277 282 284 285 286 288 289 292 298 304 305 307

Contents

10.5.3 Session Identiﬁers 10.5.4 Business Service Realization 10.6 Using Patterns, Frameworks and Libraries 10.7 Transactions 10.7.1 Pessimistic and Optimistic Concurrency 10.7.2 General Guidelines for Using Transactions with Objects 10.7.3 Transactions in Upper Layers 10.8 Handling Multiple Activities 10.8.1 Controlling Multiple Tasks 10.8.2 Controlling Multiple Threads 10.8.3 Thread Safety 10.9 Summary Further Reading Review Questions Answers to Review Questions 11 Reusable Design Patterns 11.1 Introduction 11.1.1 A Brief History of Patterns 11.1.2 Software Patterns Today 11.2 A Pattern Template 11.3 Common Design Patterns 11.3.1 Observer 11.3.2 Singleton 11.3.3 Multiton 11.3.4 Iterator 11.3.5 Factory Method and Abstract Factory 11.3.6 State 11.3.7 Facade 11.3.8 Adapter 11.3.9 Strategy and Template Method 11.3.10 Flyweight 11.3.11 Composite 11.3.12 Proxy 11.4 Using Patterns 11.5 Discovering, Combining and Adapting Patterns 11.6 Summary Further Reading

309 310 312 312 313 314 315 315 315 316 318 321 321 322 323 326 328 328 329 330 331 331 337 341 342 346 347 352 354 356 358 361 364 367 367 370 371

xiii

xiv

Contents

12 Specifying the Interfaces of Classes 12.1 Introduction 12.2 What Is a Speciﬁcation? 12.3 Formal Speciﬁcation 12.4 Informal Speciﬁcation 12.5 Dynamic Checking 12.6 Object-Oriented Speciﬁcation 12.6.1 Formal Speciﬁcation in OCL 12.6.2 Informal Speciﬁcation in Eiffel 12.7 Design by Contract 12.7.1 Contracts and Inheritance 12.7.2 Reducing Error-Checking Code 12.7.3 Enforcing the Contract 12.7.4 Application Firewalls 12.8 Informal Speciﬁcation in Java 12.8.1 Documenting a Contract using Comments 12.8.2 Checking Conditions Dynamically 12.8.3 Signaling Contract Violations using RuntimeExceptions 12.8.4 External Systems 12.8.5 Enabling and Disabling Dynamic Checks 12.9 Summary Further Reading Review Questions Answers to Review Questions

372 374 375 376 378 380 382 383 384 385 389 391 394 395 396 396 396 397 398 400 402 403 403 404

13 Continuous Testing 13.1 Introduction 13.2 Testing Terminology 13.2.1 Black-Box Testing 13.2.2 White-Box Testing 13.3 Types of Test 13.3.1 Unit Testing 13.3.2 Integration Testing 13.3.3 Alpha Testing 13.3.4 Beta Testing 13.3.5 Use Case Testing 13.3.6 Component Testing 13.3.7 Build Testing 13.3.8 Load Testing 13.3.9 Installation Testing

406 408 408 409 410 411 412 412 413 413 414 414 415 417 418

Contents

13.3.10 Acceptance Testing 13.3.11 Regression Tests 13.3.12 Documentation Tests 13.3.13 Testing for Security 13.3.14 Metrics 13.4 Automating Tests 13.5 Preparing for Testing 13.6 Testing Strategies 13.6.1 Testing During Development 13.6.2 Testing During the Testing Phase 13.6.3 Testing After Release 13.7 What to Test For 13.8 Test-Driven Development 13.9 An Example of Test-Driven Development using JUnit 13.9.1 Testing the Car Class 13.9.2 Implementing the Car Class 13.9.3 Refactoring Tests 13.9.4 Creating a Test Suite for Regression Testing 13.9.5 Testing Across Methods 13.9.6 Completing the Store Class 13.10 Summary Further Reading

418 419 419 419 420 421 422 424 424 425 426 426 430 431 433 434 435 439 441 442 445 445

A Ripple Summary

446

B iCoot Case Study B.1 Business Requirements B.1.1 Customer’s Mission Statement B.1.2 Actor List B.1.3 Use Case List B.1.4 Use Case Communication Diagrams B.1.5 Use Case Activity Diagrams B.1.6 Use Case Details B.2 System Requirements B.2.1 User Interface Sketches B.2.2 Actor List B.2.3 Use Case List B.2.4 Use Case Diagram B.2.5 Use Case Survey

450 450 450 450 451 452 452 452 456 456 460 460 460 461

xv

xvi

Contents

B.3

B.4

B.5

B.6

B.7

B.2.6 Use Case Details B.2.7 Supplementary Requirements B.2.8 Use Case Priorities Analysis B.3.1 Class Diagram B.3.2 Attributes B.3.3 Operation List B.3.4 State Machine for a Reservation B.3.5 Use Case Realization System Design B.4.1 Technology Choices B.4.2 Layer Diagram B.4.3 Layer Interaction Policy B.4.4 Packages B.4.5 Deployment Diagram B.4.6 Security Policy B.4.7 Concurrency Policy Subsystem Design B.5.1 Business Services B.5.2 ServletsLayer Class Diagram B.5.3 ServletsLayer Field List B.5.4 ServletsLayer Message List B.5.5 ServerLayer Class Diagram B.5.6 ServerLayer Field List B.5.7 ServerLayer Message List B.5.8 BusinessLayer Class Diagram B.5.9 BusinessLayer Field List B.5.10 Protocol Objects Class Diagram B.5.11 Database Schema B.5.12 User Interface Design B.5.13 Business Service Realization Class Speciﬁcation B.6.1 Server Class Speciﬁcation B.6.2 Business Logic Class Speciﬁcation Outline Test Plan B.7.1 Introduction B.7.2 The Impact of Spirals and Increments B.7.3 Testing of Non-Code Artifacts B.7.4 Code Reviews

462 465 465 466 466 467 467 469 470 476 476 477 479 479 479 482 482 483 483 483 484 484 485 486 486 487 488 492 494 494 494 508 508 509 512 512 512 513 513

Contents

B.7.5 Test-Driven Development B.7.6 Assertions B.7.7 Testing Phase B.7.8 Documentation Testing B.7.9 Build Testing B.7.10 Test Documentation and Logging B.7.11 Testing Activities by Phase B.8 Glossary

513 514 514 515 515 515 516 517

C Summary of UML Notation Used

526

Bibliography

538

Index

541

xvii

1 Introduction

T

he aim of this book is to give you a broad understanding of the processes and techniques used in object-oriented software development, the production of computer programs using object-oriented techniques. The Uniﬁed Modeling Language (UML) comes into play as the standard notation used in industry for software documentation. You may be a student at a university or on a commercial training course. Or you may be an experienced software developer, moving into object orientation for the sake of your career. In either case, this book is for you. Little prior knowledge is required on your part and there is no attempt to teach you everything there is to know. Instead, you will be guided through the essential parts of the process, so that you can do your job more effectively. Although this book is broad, it only goes as deep as the point where you would normally start writing lines of code. Describing how to write code would mean choosing a particular programming language; the decision about the most appropriate language is for you to make. You should consider this book as a generic front-end that will work for any pure, object-oriented programming language. The objectives of this ﬁrst chapter are to describe the background to this book, to give you an idea of the content and to describe how to navigate your way around.

Chapter Outline 1.1 Background 1.2 A Brief History of Programming 1.3 Methodologies 1.4 About this Book

1

4

Chapter 1

1.1 BACKGROUND These days, new software is usually object-oriented. That is, the software is written using an abstraction called an object. There is, naturally, much more to commercial software development than simply writing lines of code: there is investigation of the business requirements, analysis of the problem, design of the solution, and so on. Objects should be used at every stage of the development because they reduce the amount of information that has to be understood and improve the communication between members of the development team.

1.2 A BRIEF HISTORY OF PROGRAMMING Commercial programming has had a number of generations, of which ‘object-oriented’ is just the latest: • Machine code: Programming using binary numbers. • Assembly language: Programming using alphanumeric symbols, or mnemonics, as shorthand for machine code. Assembly language is translated into machine code by a program called an assembler. • High-level languages: Programming using languages (such as Fortran and COBOL) that have high-level constructs such as types, functions, loops and branches. High-level languages (and later generations of programming languages) are translated into machine code using a program called a compiler. • Structured programming: Programming using cleaner high-level languages (such as Pascal, Modula and Ada) that are characterized by fewer pitfalls for the programmer and more discipline in the way a program is broken down into sub-tasks and sub-systems. • Object-oriented programming: Programming using independent modules of data and functions that correspond to concepts in the problem domain, such as Customer or ScrollBar. This modularity leads to even fewer pitfalls for the programmer and encourages the reuse of code across separate programs. Good object-oriented programming languages include Java and Eiffel, because they’re well designed, pure and portable (available on many platforms). Other examples include Smalltalk, C# and, in general, any language that started life as a structured language and grew up with object-oriented extensions (C++ and various dialects of Pascal, for example). You may also have heard of functional programming and logic programming. So far they have had little commercial impact. All the generations above survive today, to a greater or lesser extent. Which one we end up using depends on the situation we walk into; personal preference; and the constraints of

About this Book

the problem that we’re trying to solve – for example, video games need every ounce of raw speed, so they’re sometimes written in assembly language.

1.3 METHODOLOGIES Around the time that structured programming was becoming popular, in the 1980s, experienced programmers began trying to describe how the entire software development process should be controlled, from mission statement through to maintenance of the ﬁnished product. This led to structured methodologies such as SSADM [Weaver et al. 02]. A methodology is a description of the steps a development team should go through in order to produce a high-quality system. A methodology also describes what should be produced (documents, diagrams, code, etc.) and what form the products should take (for example, content, icons, coding style). When object-oriented programming was catching on, in the 1990s, developers invented object-oriented methodologies, better suited to an object-oriented programming style. Early object-oriented methodologies included the Booch method [Booch 93], Objectory [Jacobson et al. 92] and OMT [Rumbaugh et al. 91]. These days, one of the market leading methodologies is the Rational Uniﬁed Process (RUP) [Jacobson et al. 99], owned by IBM (www.rational.com). Roughly speaking, RUP is a convergence of Objectory, Booch and OMT. Another methodology that is gaining in popularity is extreme programming (XP) [Beck 99], a so-called ‘agile’ methodology – in the context of software development, agile means responsive to changes in software requirements or changes in our understanding of the problem. The methodology used in this book, Ripple (summarized in Appendix 1), is a simpliﬁed one based on widely accepted theory and practice. As a result, you won’t have to learn the complexities of a full industrial methodology. Nor will you be told exactly what to do at each stage, which will allow you to be more creative as you learn.

1.4 ABOUT THIS BOOK In order to avoid confusion, there is no detailed comparison between structured and objectoriented methodologies in this book. Instead, you will be taken through object-oriented software development, as if traditional methods had never existed. You will encounter everything you need to know to start producing large amounts of good, object-oriented software (you just have to add the effort and the experience yourself). If you’re already familiar with structured techniques, you may ﬁnd yourself having to un-learn things from time to time. But don’t worry, the object-oriented approach really does

5

6

Chapter 1

work: it’s been around in the research community since 1970, in the marketplace since 1990 and it’s used every day by millions of developers. As ever, the goal of software development is code, code and more code. Whatever your particular background, if you’ve got experience programming a computer, you’ll be on safe ground. On the other hand, if you don’t have much programming experience, you’ll be pleased to learn that this book is straightforward – there’s no attempt to make your head spin with obscure jargon and magic tricks. However, before you start, you should make sure that you’re comfortable with basic computer concepts such as hardware, software and networks. At the very least, you should have written a few hundred lines of code in a high-level language.

1.4.1 Content Overview Although the writing of code won’t be covered in any detail, from time to time there will be a need to illustrate a point using code fragments. All the code fragments are written in Java [Joy et al. 00] because it is popular, pure, simple and free. The meaning of each code fragment is clearly explained in the text. If you’re not a great fan of Java, rest assured that the code fragments can easily be translated into other languages such as C#, because the unique elements of Java have been avoided. Everything you see here in Java can be accomplished in any other language that you choose. Similarly, the discussion of system design is focussed on Java technologies rather than .Net technologies – since Java and .Net provide similar facilities, consider this simply a matter of personal preference. All the pieces of Java system design that are presented can be implemented in a similar way using .Net. A few words of warning, however: for illustrative purposes, classes (such as Iterator and List) are described that don’t match those in the Java library exactly. Although you can certainly use this book to see how Java works and get familiar with most of the syntax, it is no substitute for a pure Java language text [Campione et al. 00]. In any case, it is good practice always to have the library documentation to hand when writing in Java, since no-one could keep the details of thousands of classes in their head. Another major omission is project management (issues such as planning and scheduling). Project management is omitted in order to focus on technical issues, rather than human factors. The notation used for illustrations, wherever possible, is the Uniﬁed Modeling Language (UML) [OMG 03a]. This has become the accepted standard for software diagrams. UML presentation conventions mean that some lines are thicker than others and that some characters are in bold or italics. Some of these conventions are difﬁcult to accomplish when drawing by hand (on a piece of paper or on a whiteboard, for example), however only the italicized text is really important, so the other conventions can be ignored when hand-drawing. For italicized text, alternatives will be identiﬁed at the appropriate time.

About this Book

UML doesn’t satisfy all documentation needs, so some of the documentation ideas in this book have been taken from RUP.

1.4.2 Case Studies The case study used throughout is a rental and reservation system called Coot, developed for a ﬁctitious company called Nowhere Cars. Thus, many of the examples in the text will use cars in one way or another. Using the same application area for most of the examples means that you don’t have to keep adjusting to different areas from one page to the next. To keep things simple, most of the discussion involves those parts of Coot that provide Internet facilities to customers, a cut-down system called iCoot. Because there are many concepts to illustrate, not all of which are relevant to the renting of cars, some of the examples in the text do not make it into iCoot itself (they would be more appropriate for car salespeople or car mechanics). However, every time you’re presented with a major diagram, it will be one that has been taken from the ﬁnished system. Because newcomers to object orientation often ask for full case study documentation, the ﬁnished documentation for iCoot has been included in Appendix B, for further study. The iCoot documents have been designed to be understandable, despite representing a realistic and useful piece of software. If you want to try out the techniques described in this book with fellow students or colleagues, a set of group exercises is available at www.wiley.com/go/odocherty for the Automated Quiz System (AQS). AQS is an on-line tool for taking multiple-choice quizzes. The exercises are organized according to the main chapters of this book, so that you can complete them while you read. For lecturers and instructors wishing to use this book as a course text at a University or commercial training organization, sample solutions are available for the AQS exercises, subject to registration.

1.4.3 Navigation After reading Chapter 1, you can take the straightforward approach to navigation and work through all the chapters in order. Alternatively, if you’re already comfortable with object-oriented concepts and jargon, you could skip Chapters 2, 3 and 4 and jump straight to Chapter 5. If you’re completely new to the ﬁeld of object orientation, you should read Chapter 2, but you may like to leave Chapters 3 and 4 until a second reading. If you’re not interested in an overview of Ripple and methodologies in general but would prefer to get straight down to the details, you could skip Chapter 5 and go straight on to Chapter 6. The remaining chapters follow a strict sequence, the progress of a typical software development, so jumping around is not recommended. It would be valuable to read

7

8

Chapter 1

Chapter 11 as part of the main sequence; however, since many of the issues are advanced, you may wish to leave it until a later date. You can use Appendix 1 as a reference whenever you need to check where you’ve got to or if you’re attempting the case study for yourself. Appendix B contains the ﬁnished artifacts for the iCoot case study, organized according to the progress of a typical development. Use these, along with the main development chapters, to see how the iCoot artifacts turned out. Appendix B also includes the iCoot project glossary, which was continually updated during iCoot development. Use this to see how a typical glossary might look and to look up the deﬁnition of any iCoot term. Whenever you need to remind yourself how to draw part of a UML diagram, you should refer to Appendix 3. Happy reading.

I

Setting the Scene

2

Object Concepts

T

he concepts we’ll look at in this chapter come out of object-oriented programming. Typically, programming languages are invented before the methodologies that help us to use them. You should ﬁnd that object-oriented concepts make a lot of sense. This is because object-oriented development allows you to think in real-world terms, rather than bending your mind towards the needs of a computer.

Learning Objectives Understand what we mean by a software object.

•

Understand how objects can communicate, using messages, to accomplish some task.

Understand what happens when an object is no longer needed (garbage collection).

•

Understand what we mean by a class.

•

Understand how to reuse code.

Chapter Outline 2.1 Introduction 2.2 What Is an Object? 2.3 Identical or Equal? 2.4 Depicting Objects 2.5 Encapsulation 2.6 Association and Aggregation 2.7 Graphs and Trees 2.8 Links and Navigability 2.9 Messages 2.10 Invoking an Operation 2.11 An Example Collaboration 2.12 How an Object-Oriented Program Works 2.13 Garbage Collection 2.14 Classes 2.15 What Does a Class Deﬁne? 2.16 Shared Data and Shared Operations 2.17 Types 2.18 Terminology 2.19 Reusing Code 2.20 Summary Further Reading Review Questions Answers to Activity 1 Answers to Activities Answers to Review Questions

2

12

Chapter 2

2.1 INTRODUCTION The basic concepts of the object-oriented paradigm (way of doing things) are relatively easy to understand and to apply. Alan Kay, the inventor of Smalltalk, had been working on ‘A Personal Computer for Children of all Ages’ [Kay 72] as early as 1968: as his target was children, it isn’t surprising that the basic concepts are simple. So, why all the fuss about objects? Surely developers wouldn’t change the fundamentals of software development without good reason? Some of the justiﬁcations for using objects might seem rather obscure at this early stage, especially if you haven’t much experience with the techniques that came before (structured programming and structured methodologies). The object-oriented approach was invented (or, rather, it evolved) because of the difﬁculties people were having trying to get good quality systems produced on time and within budget, especially for large systems with many people involved. Once you’ve worked your way through this book, the justiﬁcations given below should make sense and you should agree with most of them. Here then, for the record, are some of the justiﬁcations typically given for object orientation: • Objects are easier for people to understand: This is because the objects are derived from the business that we’re trying to automate, rather than being inﬂuenced too early by computer-based procedures or data storage requirements. For example, in a bank system, we program in terms of bank accounts, bank tellers and customers, instead of diving straight into account records, deposit and withdrawal procedures, and loan qualiﬁcation algorithms. • Specialists can communicate better: Over time, the software industry has constructed career ladders that newcomers are expected to climb gradually as their knowledge and experience increases. Typically, the ﬁrst rung is programmer: ﬁxing faults (bugs) in the code written by others. The second rung is senior programmer: writing the code itself. The third is designer: deciding what code needs to be written. Finally comes the role of analyst: talking to customers to discover what they need and then writing down a speciﬁcation of what the ﬁnished system must be able to do. Such a career ladder may not be a bad idea in itself. The problem comes when you realize that each specialist is expected to learn a whole new set of concepts and techniques, depicting their conclusions using notations that are tailored to their specialty. This means that there’s a big gap in understanding between the different roles, made worse by the fact that the documents are being passed down the career ladder rather than up, so we tend to have to read documents without understanding the techniques used to produce them. This can lead to ‘throw it over the wall’ syndrome: the analyst produces a large amount of documentation, throws it over the wall to the designer and walks away. The designer, after weeks of effort, produces even more documentation, using completely

What Is an Object?

different techniques, and throws it over the wall to the programmers. The programmers then start all over again . . . With the object-oriented approach, everyone is dealing with the same concepts and notations. Moreover, there are generally fewer concepts and fewer notations to deal with in the ﬁrst place. • Data and processes are not artiﬁcially separated: In traditional methods, the data that needs to be stored is separated early on from the algorithms that operate on that data and they are then developed independently. This can result in the data being in inconvenient formats or inconvenient locations, with respect to the processes that need access. With objectoriented development, data and processes are kept together in small, easy-to-manage packages; data is never separated from the algorithms. We also end up with less complex code that is less sensitive to changes in customer requirements. • Code can be reused more easily: With the traditional approach, we start with the problem that needs to be solved and allow that problem to drive the entire development. We end up with a monolithic solution to today’s problem. But tomorrow always brings a different problem to solve; no matter how close the new problem is to the last one we dealt with, we’re unlikely to be able to break open our monolithic system and make it ﬁt – we hamper ourselves by allowing a single problem to inﬂuence every part of the system. With object-oriented development, we’re constantly looking for objects that would be useful in similar systems. Even when a new system has minor differences, we’re much more likely to be able to change our existing code to ﬁt, because objects are like the pieces in a jigsaw puzzle: if one piece is changed, it might affect a few pieces next to it, but the rest of the puzzle will remain intact. When we build an object-oriented system, we try to ﬁnd existing objects (written by us, by our colleagues or by third parties), before we consider writing any code ourselves. As one sage put it, ‘object-oriented programming is about not writing code’. • Object orientation is mature and well proven: This is not a new fad. The programming concepts emerged in the late 1960s while the methodologies have been around for at least a decade. Applying objects in such areas as software, databases and networks is now well understood. Once you’ve read the whole of this book, try reviewing this list to see if you fully understand, and agree with, the justiﬁcations.

2.2 WHAT IS AN OBJECT? An object is a thing, an entity, a noun, something you can pick up or kick, anything you can imagine that has its own identity. Some objects are living, some aren’t. Examples from

13

14

Chapter 2

the real world include a car, a person, a house, a table, a dog, a pot plant, a check book or a raincoat. All objects have attributes: for example, a car has a manufacturer, a model number, a color and a price; a dog has a breed, an age, a color and a favorite toy. Objects also have behavior: a car can move from one place to another and a dog can bark. In object-oriented software, real world objects migrate into the code. In programming terms, our objects become stand-alone modules with their own knowledge and behavior (or, if you prefer, their own data and processes). It’s common to think of a software object as a robot, an animal, or a little person: each object has certain knowledge, in the form of attributes, and it knows how to perform certain operations for the beneﬁt of the rest of the program. For example, a person object might know its title, ﬁrst name, last name, date of birth and address; it would be able to change its name, move to a new address, tell us how old it is, and so on. By concentrating on the characteristics of a person when we write the code for the Person object, we can put the rest of the system out of our mind – this makes our programming simpler than it would otherwise be (it also helps that we have a real-world concept to get us started). If we decide later that our Person needs to know its height, we can add that extra knowledge (and associated behavior) directly to the Person code. Only the code in the rest of the system that needs to use the height attribute has to be modiﬁed; all other code would remain unchanged. Simplicity and localization of change are important characteristics of object-oriented software. It’s easy enough to think of a living thing as some kind of robot. It is a little strange, though, when we try to think of a lifeless object as having behavior. We wouldn’t normally consider a video capable of changing its price or giving itself a new advert. However, when it comes to object-oriented software, that’s exactly what we need to do. The key is that if the video didn’t do those jobs, some other part of the system would have to. This would lead to video characteristics leaking into other parts of the code, so we’d start to lose the simplicity and locality that we crave (we would be going back to ‘the old way of doing things’). Don’t be put off by the anthropomorphism (the assignment of human characteristics to inanimate objects or animals) that’s common in object-oriented development, imagining software objects as little people. Figure 2.1 shows some real-world objects that would make good software objects. Can you think of any others? Can you think of anything that wouldn’t make a good object? Well, that’s a trick question really: ‘thing’ at the end of ‘anything’ suggests that the answer has to be no. It turns out that almost anything would make a good object in some context or another. A bad object is one that merges several concepts, for example, a bank account object that has been polluted with the knowledge and behavior that belongs to a bank clerk. Always keep in mind that separate concepts in the real world suggest separate concepts in the program.

What Is an Object?

aPerson

aBankAccount

aNumber

aDate

aCat

Figure 2.1: Objects in the real world

Before we go any further, it’s important to note that we’re not trying to simulate the real world, that would be far too difﬁcult. we’re simply trying to make sure that our software is inﬂuenced by real-world concepts, so that it is easier to produce and easier to change. The needs of the system and the needs of computers are important inﬂuences too. Some developers are uncomfortable with the close proximity of the real world and software; however, an object-oriented system developed for a hospital that didn’t include some kind of Patient object would be of little use. You shouldn’t think that it’s possible to code the ideal Person object or any other kind of perfect object. There are simply too many characteristics and capabilities that can be applied to real-world objects – if we tried to capture them all, we’d never get as far as coding the second object in our system. Most aspects of a real-world object are not needed for a typical program, especially since software systems tend to address a single problem area. For example, a bank system will be interested in the age and salary of a customer, but not their shoe size or favorite color. It’s quite reasonable to code an object that is useful for many systems, especially for well understood areas of programming: for example, all systems with a user interface are likely to

15

16

Chapter 2

be able to make use of the same ‘scrollable list’ object. The trick is to start by considering the business you’re dealing with, to ask yourself ‘If I worked in this business area, what would a ‘‘person’’ mean to me: a customer, an employee, a patient, or what?’ All good software developers start by modeling the business. A model is a representation of a problem domain or a proposed solution that allows us to talk, or reason, about the real thing. This allows us to increase our understanding and avoid potential pitfalls. Think of an architect’s model of a new concert hall: it allows the architects to say ‘This is what the ﬁnished concert hall will look like’ and it helps them to come up with new ideas, such as ‘I think we’re going to need a steeper roof’. A model allows us to learn a lot without actually building anything. Much of software development involves creating and reﬁning models, rather than cutting lines of code.

Implementation Point 1 Let’s consider how we can create a new object in an object-oriented programming language. Pure object-oriented languages typically provide a creation expression. Here’s how one looks in Java: new Person("Sue Smith")

The effect of this expression is to create space for a new Person object and pass it the string ‘‘Sue Smith’’, so that it can initialize itself (presumably, in this case, the object would record its name). Once we’ve created an object, we can put it somewhere where we can ﬁnd it later, by assigning a name to it, as in: aPerson = new Person("Sue Smith");

Now, whenever we write down aPerson, we will be referring to the object that we just created.

2.3 IDENTICAL OR EQUAL? Objects have their own independent existence. Consider holding a blue pen in your left hand and a blue pen in your right hand. You’re holding separate pens: they exist independently and each one has its own identity. But the pens can have similar attributes: blue ink, half full, same manufacturer, same model, etc. In terms of their attributes, the pens are interchangeable – if you wrote something down on a piece of paper, no-one would be able to

Identical or Equal?

tell which pen you’d used (unless they saw you do it). The pens are equal but not identical. This is an important distinction in software, as well as in real life. To take another example, consider the situation illustrated in Figure 2.2. In Acacia Avenue there are two families, the Smiths and the Joneses: the Smiths live at number 4 and the Joneses live at number 7. The families have similar tastes in lawn mowers; so similar, in fact, that they both own a GrassMaster 75, purchased on the ﬁrst day the new model became available. The lawn mowers are so similar that if someone switched them round overnight, the Smiths and the Joneses wouldn’t notice. As well as a lawn mower, the Smiths have a cat called Tom. Tom is a round friendly cat, three years old, whose favorite pastime is chasing mice around the local gardens. The Joneses also have a cat, called Tiddles. Tiddles is a round, friendly cat, three years old, whose

Figure 2.2: Identical or equal?

17

18

Chapter 2

favorite pastime is chasing balls of wool. Anyone visiting the Smiths and the Joneses would notice a remarkable similarity between Tom and Tiddles, not surprising when you realize that Tom and Tiddles are the same cat, round on account of being fed so often. In this situation, we have two lawn mowers and one cat. Although the lawn mowers have a separate identity, recorded on the serial-number plates riveted to their bodies, they are equal, because they have the same attributes. The cat also has an identity, it may even have a name for itself (‘Me’ or ‘Hfrrr’, perhaps). The difference between the cat and the lawn mowers is that the cat is shared and the lawn mowers aren’t. It’s rarely necessary for any person, thing or animal to be concerned with its own identity: the cat doesn’t think about whether it’s separate from other cats; a lawn mower doesn’t need to know which lawn mower it is in order to mow the lawn. The families don’t need to know that Hfrrr is fed twice as much as other cats, as long as he purrs in their laps from time to time; nor do the families need to know that the lawn mower in their shed is the one they actually purchased, as long as it’s there when they need it. Generally speaking, in an object-oriented system, if we use one software object to represent each real world object, we won’t go far wrong. So, in the example that we’ve just seen, we would expect to ﬁnd in our system two lawn mower objects and a single cat object. We may sometimes share objects. We may sometimes swap equal objects. But we rarely need to worry about identity: we just tell an object what to do and it uses its knowledge and capabilities to carry out the request.

Implementation Point 2 Object-oriented programming languages allow us to test whether objects are identical or equal if necessary. The following fragment of Java code does just that. There are two tests in it: if (tom == tiddles) is a test of identity that evaluates to true if tom and tiddles both point to the same object; tom.equals(tiddles) is an instruction that asks tom to perform the equals operation, with tiddles as a parameter – it evaluates to true if tom and tiddles are separate but equivalent to each other. if (tom == tiddles) { result = "The Smiths and the Joneses share one cat"; } else if (tom.equals(tiddles)) { result = "The Smiths and Joneses have equivalent cats"; } else { result = "The Smiths and the Joneses have different cats"; }

Depicting Objects

2.4 DEPICTING OBJECTS Once we’ve decided to work with objects, we need some way of showing them on a diagram so that we can describe them and think about them. Figure 2.3 shows how we can draw an object. The notation used here is a UML object diagram – you’ll see more details of UML and its history as you go through this book; for now, the notation will be introduced by example.

anObject attribute1 attribute2 operation1() operation2() operation3()

standard notation but in non-standard location

Figure 2.3: An object

The three parts of the box show the object’s name (which is underlined), its attributes (its knowledge) and its operations (its behavior). (Showing operations on an object diagram is not actually legal UML, but it suits our purposes for now.) The parentheses next to the operation names indicate any parameters that are required: even though none of the operations shown here have parameters, it’s a good idea to keep the parentheses so that operation names stand out from attribute names (this becomes even more important when we omit one or more parts of the box). The attributes are hidden inside the object; the only way to access them is via the operations. This is reasonable from analogy with the real world: most of us are much more interested in the fact that a television has a ‘change channel’ operation than we are in the fancy electronics inside the box that actually perform the change. Let’s examine a coffee machine object. First, we should decide what operations a coffee machine needs: • • • •

display drinks select drink accept money dispense drink

Next, we should think about what the coffee machine needs to know in order to perform these operations:

19

20

Chapter 2

• available drinks • drink prices • drink recipes Having designed our coffee machine object, we can record our ﬁndings on an object diagram, like the one shown in Figure 2.4.

aCoffeeMachine drinkPrices availableDrinks drinkRecipes displayDrinks() selectDrink() dispenseDrink() acceptMoney()

Figure 2.4: A coffee machine object

2.5 ENCAPSULATION Encapsulation refers to an object hiding its attributes behind its operations (it seals the attributes in a capsule, with operations on the edge). Hidden attributes are said to be private. Some programming languages (for example, Smalltalk) automatically make attributes private and some languages (for example, Java) leave it to the programmer. Encapsulation is one of the ways that a programming language protects programmers from themselves: if programmers could bypass the operations, they would become dependent on the attributes that were being used to represent the object’s knowledge. It would then be much harder to change the internal representation of the object in the future, because we’d have to ﬁnd all the pieces of code that access the attributes directly and change those too. Without encapsulation we would lose simplicity and locality. As an example of why encapsulation is a good idea, consider an object representing a circle. A circle would be likely to have operations allowing us to discover its radius, diameter, area and perimeter. What attributes would we need to store in order to support this behavior? Well, we could store the radius and calculate the other attributes on demand. Or, we could store the diameter and calculate the other attributes from that. In fact, we could store any one of the four attributes and calculate the other three on demand. (Our choice may depend on personal preference or it may depend on predicting how the circle will normally be used.)

Association and Aggregation

Let’s say we choose to store the diameter. Any programmer who was allowed to access the diameter might do so, rather than going via the ‘get diameter’ operation. If, for a later version of our software, we decided that we wanted to store the radius instead, we would have to ﬁnd all the pieces of code in the system that used direct access to the diameter, so that we could correct them (and we might introduce faults along the way). With encapsulation, there’s no problem. Another way to think of encapsulation is to imagine that objects are courteous to one another. If you wanted to borrow some money from a colleague to buy food in the staff canteen, you wouldn’t grab their wallet and look through it to see if they had enough cash. Instead, you would ask them whether they could lend you some money and they would look in their own wallet.

2.6 ASSOCIATION AND AGGREGATION No object is an island. All objects are connected to other objects, directly or indirectly, strongly or loosely. By connecting objects, we make them more powerful. Connections allow us to navigate around to ﬁnd extra information and behavior. For example, if we were processing a Customer object representing Freda Bloggs and we wanted to send Freda a letter, we would need to know that Freda lives at 42 Acer Road. We would expect the address information to be stored in some kind of Address object, so we would look for a connection from the Customer to the Address, to ﬁnd out where to send the letter. When we’re modeling with objects, we can connect them in two principal ways: association or aggregation. It’s sometimes hard to spot the difference between the two, but here are some ideas. • Association is a weak form of connection: the objects may be part of a group, or family, of objects but they’re not completely dependent on each other. For example, consider a car, a driver, a passenger and another passenger. When the driver and the two passengers are in the car, they’re associated: they all go in the same direction, they occupy the same volume in space, and so on. But the association is loose: the driver can drop off one of the passengers to go their separate way, so that the passenger is no longer associated with the other objects. Figure 2.5 shows how we can draw an association on an object diagram – the attributes and operations have been omitted here in order to emphasize the structure. • Aggregation means putting objects together to make a bigger object. Manufactured items usually form aggregations: for example, a microwave is made up of a cabinet, a door, an indicator panel, buttons, a motor, a glass plate, a magnetron, and so on. Aggregations usually form a part–whole hierarchy. Aggregation implies close dependency, at least of

21

22

Chapter 2

aPassenger

aCar

aDriver anotherPassenger

Figure 2.5: Association

the whole to the part; for example, a magnetron is still a magnetron if you take it out of its microwave, but the microwave would be useless without the magnetron, because it wouldn’t be able to cook anything. Figure 2.6 shows how we can draw a house as an aggregation: in order to emphasize the difference between this kind of connection and an association, we place a white diamond on the ‘whole’ end.

diningRoom

livingRoom firstFloor

kitchen aHouse mainBedroom

secondFloor kidsBedroom

bathroom

Figure 2.6: Aggregation

Graphs and Trees

As suggested, the distinction between association and aggregation can be subtle. The ‘What happens if you remove one of the objects?’ test can be helpful, but it doesn’t always solve the problem: hard thinking and experience are often needed. We often need to choose between aggregation and association, because the choice can affect the way we design our software. Here are some examples: • Friends: We would expect friends to be an association: there is no sense in which we could put the friends together to make a larger friend; friends come and go over time. • Components in a television set: This is one of the easier ones to reason about because it’s the classic part–whole hierarchy: you put the buttons and the knobs together to make the control panel; you put the glass screen, the electron gun and the magnetic coils together to make the tube; once the small parts have been assembled, then assembled into bigger components, you put everything into the cabinet and screw the back on. The end users see a single television object: if one of the components fails, they probably don’t think they have a television set anymore, just a useless heap of junk. • Books on a bookshelf: A bookshelf doesn’t need books to be a bookshelf, it’s just a place to put the books that we own. Conversely, when a book is on a bookshelf, it is certainly associated with it (if you move the bookshelf, the book moves too; if the bookshelf collapses, the book falls). This is a classic association. • Windows in an ofﬁce block: The windows are part of the ofﬁce block. Although we could remove a broken window, leaving the ofﬁce one window short, we would expect a replacement to be provided soon afterwards. This is probably aggregation.

ACTIVITY 1

Now it’s your turn. Which of the following examples are association and which are aggregation? 1. Houses on a street. 2. Pages in a book. 3. Notes in a symphony. 4. Components in a home entertainment system (television, VCR, tape deck, ampliﬁer, games console).

2.7 GRAPHS AND TREES As well as associations and aggregations, you may hear about trees or graphs of objects. A tree is another name for a hierarchy. If we redraw the object diagram from Figure 2.6, as we have in Figure 2.7, you can see why aggregation is often referred to as a tree (it doesn’t have a trunk, but it’s close enough). For reasons best known to themselves, programmers usually draw trees upside down, as shown in Figure 2.8.

23

24

Chapter 2

kidsBedroom

livingRoom

diningRoom

kitchen

bathroom

mainBedroom

secondFloor

firstFloor

aHouse

Figure 2.7: Aggregation as a tree

aHouse

secondFloor

firstFloor

diningRoom

kitchen

bathroom

mainBedroom

kidsBedroom

livingRoom

Figure 2.8: An upside-down tree

Links and Navigability

A graph is an arbitrary set of connections between a group of objects. Objects in an association often form a graph, as in the car example in Figure 2.5. Another example, with more interesting connections, is the underground train system in London. Figure 2.9 represents part of the London Underground system: in this case it’s possible to get from any node (station) to any other node, usually via several routes.

Euston

Warren Street

King's Cross St. Pancras

Goodge Street

Oxford Circus

Tottenham Court Road

Leicester Square Picadilly Circus Charing Cross

Embankment

Figure 2.9: An object graph showing connections between London Underground stations

A tree is a special case of a graph: each node in a tree has only one parent node, but it can have any number of children. (A parent contains the children, in a part–whole sense.) This corresponds well to the difference between aggregation and association: any group of connected objects forms an association, but only associations which have the right inter-dependencies and the right structure qualify as trees.

2.8 LINKS AND NAVIGABILITY The connections that you’ve seen on object diagrams until now are called links. If we want to show that one object knows where the other one is, we can add an arrowhead, as shown in Figure 2.10. This shows a Customer linked to an Address and a String. (A string is a staple value in programming that comprises a sequence of characters.)

25

26

Chapter 2

anAddress address

aCustomer

house = 10 street county postcode = "TL5 1OR"

name aString

Figure 2.10: Navigable links

Each link can be thought of as an attribute: the label, or role, indicates the attribute’s name. Thus, we might say that aCustomer has an attribute, called address, that links it to an Address object, and another attribute, called name, that links it to the String representing its name. The arrowhead indicates navigability, knowing where the other object is. Because there’s no arrowhead on the customer end, the implication is that String doesn’t know that it’s associated with aCustomer. Navigable links often end up as pointers in object-oriented programs. (A pointer is the address of an object in memory, so that we can ﬁnd it when we need to.) The links in Figure 2.10 are more detailed than the connections you’ve seen until now (which didn’t have any arrow-heads). One of the advantages of an object diagram is that it allows us to show what’s going on in our model at an arbitrary level of detail – this can increase our understanding and make us more conﬁdent that we’re on the right track. Simple values are shown as attributes; important objects are shown as linked boxes; intermediate values are shown as attributes or linked boxes, as the need arises. Figure 2.10 shows other information too, information that you probably understood and accepted without too much trouble: names have been given to the linked objects and to the attributes; some literal values are also shown – the number 10 and the string TL5 1OR, for example. The naming convention used here for objects, attributes and roles is a common one: use one or two descriptive words and run them together, capitalizing each word after the ﬁrst. As for the literal values, we all know how to write down numbers and putting characters inside double quotes shouldn’t be any surprise either. In some places, objects have been expanded while, in other places, they haven’t. For example, the name attribute on aCustomer has been shown as a separate object whereas the street and county on anAddress haven’t even been given a value. The key in all diagrams is to show as much detail as needed to achieve our goal. Don’t let anyone suggest that a diagram is wrong just because they would have drawn it differently. In general, as we go through development, we will have to deal with more and

Messages

more information, but we rarely show everything in one place (if we did, things would get cluttered and tedious). A ﬁnal few words about values: although everything can be modeled as an object, for trivial values we may not bother. For example, the number 10 can be thought of as an object: it has internal data to represent its ten-ness and it has operations such as ‘add another number’ and ‘multiply by another number’. However, in many object-oriented programming languages, simple values such as numbers are treated differently: you can only use them as attribute values; they have no identity; you can’t decompose them.

2.9 MESSAGES Every object is connected to at least one other object: an isolated object wouldn’t be much use to anyone. Once objects are connected, they can collaborate, to perform more complex tasks than they could on their own. Objects collaborate by sending messages to each other, as shown in Figure 2.11. The message is shown next to a solid arrow indicating the direction in which the message is being sent; the reply is shown next to a ‘tadpole’ that indicates the movement of data. message() receiver

sender reply non-standard notation

Figure 2.11: Collaboration using messages

Figure 2.11 is a UML communication diagram. Communication diagrams look rather like object diagrams, except that the links have no direction and the object names are not underlined. Ofﬁcially, there is no way to show replies on a communication diagram, so a long-standing convention, the tadpole, has been used instead. Ideally, we would also show sequence numbers, but they’ve been omitted here, because the UML numbering scheme is rather involved. Some example messages are: ‘What’s the time?’, ‘Start the engine’ and ‘What is your name?’, as shown in Figure 2.12. As you can see, the receiving object may or may not need to provide a reply: we would expect replies to ‘What’s the time?’ and ‘What is your name?’, but not to ‘Start the engine’. As mentioned earlier, objects are courteous: when an object receives a message, it carries out the request without question. This way, the sending object doesn’t need to cope with

27

28

Chapter 2

What's the time? aClock 14:40

What is your name? aDriver

Start the engine

"Liz"

aCar

Figure 2.12: Some example messages

the possibility of a message being refused. In practice, some requests can’t be carried out, despite the best intentions of the receiver. Consider the reasons why requests can fail, listed in Table 2.1. Sometimes, we can’t allow things to fail: we would be rather disappointed if a ﬂy-by-wire plane crashed because of a software fault. Ensuring success under such circumstances is a specialist area, software reliability. Just to give you an idea, here’s one strategy for reliability: install three computers on the plane and get them to vote on what to do next; if one computer says ‘ﬂy to the left’ but the other two say ‘ﬂy to the right’, the plane ﬂies to the right. Now that you’ve seen why messages can fail, you should put the problem out of your mind for the time being and assume that messages always succeed. (This is actually good practice, as explained in Chapter 12.)

2.10 INVOKING AN OPERATION When a software object receives a message, it executes some code. As you’ve probably guessed, each piece of code is an operation. To put it another way, a message invokes an operation. In UML, we can show a message being sent from the sender to the receiver, or we can show the operation on the receiver, or we can do both. As well as replies, messages can have parameters (also called arguments). A parameter is an object or a simple value that the receiver needs in order to fulﬁll the request. We might, for example, send a person object the message ‘What is your height in meters’ one minute and ‘What is your height in feet and inches’ the next: in this case, ‘What is your height?’ would be the message, while ‘meters’ and ‘feet and inches’ would be the parameters. Parameters are shown in parentheses, after the message, as in getHeight(meters) or getHeight(aUnit). If you have several parameters, you can separate them with commas. We also need a way of specifying which object should receive a message – here’s how to do it in Java, using a full stop to separate the receiver and the message: aPerson.getHeight(aUnit)

Invoking an Operation

Table 2.1: Reasons for request failure Problem The sender should not have been allowed to send the message.

Example Sending a ﬂy message to a penguin.

Solution The compiler should spot most of these mistakes and we expect to ﬁnd the rest during testing or maintenance.

The sender makes a mistake.

Telling a microwave to start cooking when there’s no food inside.

The compiler can help but mostly we rely on good design, good programming, testing and maintenance.

The receiver makes a mistake.

Assuming that 2 + 2 = 5.

The compiler can help but mostly we rely on good design, good programming, testing and maintenance.

The receiver encounters a predictable but rare problem.

Telling a lift to ‘go up’ when there are too many people inside.

Exception handling uses the programming language’s facilities to separate normal from abnormal activity.

The computer fails to do what it’s supposed to do.

Knocking the computer off the desk; a cosmic ray blasting through the central processor; changing an internal bit from 1 to 0; an operating system fault; . . .

The software developer can’t do much apart from failing gracefully by reporting the problem to a user interface or writing to a log ﬁle.

Human error.

Removing a diskette while an object is writing information to it.

Exception handling uses the programming language’s facilities to separate normal from abnormal activity.

Sometimes, you’ll ﬁnd yourself wondering whether a message that you’re designing should get the object to do something, or retrieve some information from the object, or some combination of the two. A good guideline for message styles, one that helps us to avoid many difﬁculties, is ‘A message should be a question or a command, but not both’. A question message asks the object for some information, so it always has a reply. A question should not alter the attributes of the object (or of any object that it’s connected to). Examples of question messages are ‘What loaves do you have?’ or ‘What’s the time’. We wouldn’t expect more loaves to appear under the counter just because we’d asked the question; similarly, we would be surprised (ignoring Quantum theory) if the time inside a clock changed just because we had looked at it.

29

30

Chapter 2

A command message tells an object to do something – this time the object doesn’t need to provide a reply. Examples of commands include telling a bank account to ‘Withdraw 100 Euros’ and telling a microwave to ‘Switch off ’. If we’ve issued a reasonable command, we expect the object to go ahead and do it, so we don’t need any information back. A command alters the receiving object or some object that it’s connected to. Messages that are both questions and commands can be useful, but they’re an advanced technique, one that doesn’t warrant an example.

2.11 AN EXAMPLE COLLABORATION To solidify the concept of collaboration let’s look at a bigger example, ﬁrst from a human perspective, then from a software perspective. The example is ‘Buying a loaf of bread from a baker’s shop’. The human version is this: A customer walks into a baker’s shop and asks the baker what kind of loaves she has for sale. The baker looks under the counter and tells the customer that she has two white loaves and one wholemeal loaf. The customer says that he would like to buy the wholemeal loaf. Now the business transaction takes place: the baker wraps the loaf and offers it up with a request for payment; the customer gives the baker some money; the baker gives the customer some change. The customer leaves, satisﬁed. We can show this collaboration on a communication diagram such as the one in Figure 2.13. For simplicity, the bits about the customer entering and leaving the shop have been ignored; also, message directions have been shown alongside each message, to make things more compact. We could code this kind of collaboration pretty much directly in a pure object-oriented language. However, most object-oriented designers wouldn’t do it that way. If we consider the customer and the baker for a moment, the main problem is that we have a complicated two-way interaction – programming computers is hard enough without adding this kind of real-world complexity. Also, we have a customer that depends on the interface of the baker and a baker that depends on the interface of the customer – changing one object would mean changing the other object too, a maintenance nightmare. Contrast the design problems of our baker and customer objects with the way the baker might interact with the counter: the baker would send a message to the counter and receive a reply, but the counter wouldn’t send any messages to the baker – the counter is a passive object that sits there waiting to be used. The counter still does its job, and the baker still gets what she needs, but the interaction is one-way, making it less complex and easier to change. We call this style of interaction client–supplier: the baker is the client and the counter is supplying the services. A further advantage of supplier-style objects is that they’re more

An Example Collaboration

What do you have? Two white, one wholemeal I'd like a wholemeal I'd like some money customer

baker Here's some money Here's the loaf Here's your change

Figure 2.13: Buying a loaf of bread

getAvailableLoaves() 2 white, 1 wholemeal getPriceOf(wholemeal) €1.75 customer

baker buy(wholemeal) wholemeal pay(€2.00) €0.25

Figure 2.14: Buying a loaf, client–supplier style

likely to be useful in other contexts, because they’re independent of the client – they’re more reusable, one of the major goals of object-oriented development. We can usually transform a two-way collaboration into a client and supplier interaction with a little thought and a little practice. To help us with this, there are two mechanisms that were not used in Figure 2.13: message replies and message parameters. Figure 2.14 shows the interactions between the customer and the baker as a pure client–supplier implementation.

31

32

Chapter 2

Client–supplier is not the only way to go, but it is certainly the most common style, and it works well in most cases.

2.12 HOW AN OBJECT-ORIENTED PROGRAM WORKS An object-oriented program works by creating objects, connecting them together and getting them to collaborate by sending messages to each other. But who gets the ball rolling? Who creates the ﬁrst object and who sends the ﬁrst message? To solve this problem, all objectoriented programs have an entry point. For example, Java expects to ﬁnd an operation called main on an object that the user names when they launch the program. All the instructions in the main operation are executed, one after the other, and the program stops when main has ﬁnished. Every instruction in main can create an object, connect objects together or send a message to an object. When an object sends a message, the object that receives the message executes an operation. This operation can also create an object, connect objects together or send a message to an object. Thus, you should appreciate that this mechanism allows us to do anything we might want to do. Figure 2.15 shows an object-oriented program in action. Typically, there is not much code in a main operation – most of the behavior is inside operations on the other objects. As shown in Figure 2.15, it’s quite valid for an object to send a message to itself: a human equivalent would be asking yourself a question such as ‘What did I do yesterday?’ The idea of a main operation applies not just to programs executed from a console – it works equally well for more exotic programs such as graphical user interfaces (GUIs), Web servers and servlets. Here are some hints as to how these work: • The main operation for a user interface creates the top-level window and tells it to show itself. • A Web server’s main operation has an inﬁnite loop telling a socket object to listen for incoming requests on some port. • A servlet is an object hosted by a Web server that receives requests passed in from Web browsers – again, the Web server has the main operation.

2.13 GARBAGE COLLECTION Let’s consider what happens when an object is no longer needed by the program that created it. This may seem like a trivial issue, but remember that objects in a program don’t come

Garbage Collection

do Su bT as k( ) object1 askForHelp()

object2 data

main()

askForHelp() doSubTask()

as kF o

rM

or eH

el p(

)

m ai n(

)

data

object3 data askForMoreHelp()

Figure 2.15: An object-oriented program at run time

for free: each new object occupies a small area of the computer’s memory; as the program runs, it tends to create more and more objects, reducing the memory available to run other programs. If we didn’t reclaim objects after we had ﬁnished with them, our computer might run out of memory unnecessarily. (The memory used by a program is usually returned to the computer when the program ﬁnishes, but we might have several programs running simultaneously: some of them might be running for days, weeks or years.) It’s bad programming practice to allow more and more objects to be created by our program without taking steps to clean them up at the end of their useful life. Traditionally, programmers have had to decide for themselves when the last connection to an object was about to be removed so that they could explicitly delete or free the object’s memory. (Structured languages don’t have objects, but they do have records, structures and arrays that might need to be freed.) Keeping track of object lifetimes is complicated. It’s very easy for a programmer to forget about some of their unused objects, causing the program to keep growing – a fault called a memory leak. Languages such as Java have popularized the idea of a program that reclaims its objects automatically, without the programmer having to do anything. The basic idea is that

33

34

Chapter 2

every program has an assistant called a garbage collector, wandering around, looking for unconnected objects and sweeping them away. Sounds like magic? Well, not really. It is common these days for every program to have a run-time system – a piece of software that’s always present underneath the code that we write ourselves. It’s the run-time system that performs housekeeping tasks, such as garbage collection. Without going into any detail about how a garbage collector might work, it’s enough to know that the garbage collector is prepared to delete any object that can’t be reached, directly or indirectly, from any name that is active within the program. Any object that can’t be reached can never be sent a message; if it can never be sent a message, it can never answer a question or execute a command; therefore, it must be garbage. Pure object-oriented languages – such as Smalltalk, Java and Eiffel – tend to have garbage collectors. Hybrid object-oriented languages – such as Object Pascal – do sometimes have garbage collectors, although the fact that the languages themselves are over-complicated still means that they should be avoided if possible. C++ has no garbage collector; instead, programmers have to remember to use ‘smart pointers’, which delete an object when it appears that the last reference has gone.

2.14 CLASSES A class encapsulates characteristics common to a group of objects. There are a number of ways you could think of a class – some of them are illustrated in Figure 2.16. Putting this picture into words: • • • •

A factory manufactures objects according to some blueprint. A set speciﬁes what features its member objects will have. A template allows us to produce any number of objects of a given shape. A dictionary deﬁnition describes an object as precisely as possible.

Figure 2.17 shows some example classes. In UML, classes are drawn as boxes on a class diagram. So that we can easily tell the difference between classes and objects, class names (on class diagrams) are not underlined, while object names (on object diagrams) are. Classes and objects are rarely mixed on the same diagram: it turns out that we can do most of our modeling in terms of classes, reserving object diagrams for illustration and veriﬁcation purposes. Object-oriented programmers are often heard to say ‘Every object is an instance of a class’, hence the use of the term instance as a synonym for object. By convention, class names start with a capital letter. On a class diagram, they’re shown in bold, although, admittedly, that may be tricky when drawing by hand. In an object-oriented context, class names tend to be short and in the singular.

Classes

Dogs

factory

template

category or set

description or definition

Figure 2.16: Different ways of thinking about a class

Classiﬁcation – grouping things into classes – is something that humans are rather good at. We start doing it from the age of 12–18 months – ‘toy’, ‘food’, ‘girl’, ‘boy’, ‘doggy’ – so it’s nice to see it in programming, which is not otherwise known for its accessibility or its closeness to nature. Object-oriented software development is meant to be natural: close to the real world and close to the way that we reason about it. Since, in the real world, classes are the next major step after objects, that should be justiﬁcation enough for introducing them into our programs. Another justiﬁcation for having classes, from a software point of view, is that they allow us to share the deﬁnition of elements between related objects, so that we don’t have to repeat ourselves. Looking at Figure 2.17 again, can you spot any similarity between the classes? You would probably conclude that they’re all kinds of vehicle: some work on water, some on land and so on. By drawing lines to show how the classes are related, we come up with the picture shown in Figure 2.18.

35

36

Chapter 2

Bicycle JetPlane RowingBoat

Truck

Train

Car

Ferry Glider

Yacht

Figure 2.17: Some example classes

Train

Bicycle

Truck RowingBoat

Car Ferry

JetPlane Yacht Glider

Figure 2.18: Groups of classes

This picture shows a hierarchy of classes, a point that should be obvious when the relationships are redrawn using the proper UML notation, as in Figure 2.19. Here, a white arrow-head is used to point from each detailed concept to the less detailed one on its left. Just like the aggregation hierarchy we saw earlier, with its common name ‘part–whole hierarchy’, there are a few common names for this kind of hierarchy too: • Inheritance: Trains inherit the characteristics of land vehicles. • Generalization/specialization: A train is more specialized than a land vehicle; a land vehicle is more generalized than a train.

What Does a Class Deﬁne?

Train Truck Land Vehicle Bicycle Car Air Vehicle

Vehicle

JetPlane Glider RowingBoat

Sea Vehicle

Ferry Yacht

Figure 2.19: A class hierarchy

• Parent/child: LandVehicle is the parent of Train; Train is a child of LandVehicle. • Superclass/subclass: LandVehicle is the superclass of Train; Train is a subclass of LandVehicle. • Base/derived: LandVehicle is the base from which Train is derived. In this book, the terms base and derived won’t be used.

2.15 WHAT DOES A CLASS DEFINE? Object-oriented developers use classes to describe the programming elements that particular kinds of object will have. Without classes, we would have to add these elements to every individual object. For illustration, Figure 2.20 shows a complete class written in Java. Although this book is not about the details of Java as such, this example is simple enough for us to see in it the six essential elements of a class (see Table 2.2). New objects are created by the Actor operation; this is a special operation called a constructor that is only used when an instance of the class is created. In Java, we create an object using an expression of the form: new Actor("Charlie Chaplin");

In this case, the expression would result in a new instance of Actor with a stage name of ‘‘Charlie Chaplin’’ and a name of ‘‘’’. Operations such as getName() and setName() are known as getter and setter operations, because they get and set pieces of information.

37

38

Chapter 2

1 2 3

// An actor with "name" and "stage name" attributes public class Actor {

4 5 6 7

// Fields private String name, stageName;

8 9 10 11

public Actor(String sn) { name = ""; stageName = sn; }

// Create a new actor with the given stage name

12 13 14 15

// Get the name public String getName() { return name;

16 17 18 19

} // Set the name public void setName(String n) {

20 21 22 23

name = n; } // Get the stage name

24 25 26 27

public String getStageName() { return stageName; }

28 29 30 31

// Set the stage name public void setStageName(String sn) { stageName = sn; }

32 33 34 35

// Reply a summary of this actor’s attributes, as a string public String toString() { return "I am known as " + getStageName() +

36 37 38

", but my real name is " + getName(); } }

Figure 2.20: A simple Java class

Shared Data and Shared Operations

Table 2.2: Information deﬁned by a class Element Class name

Purpose Referring to the class elsewhere in our code.

Example in Figure 2.20 line 2

Fields

Describing the information stored by this kind of object.

name

Constructors

Controlling initialization of the objects.

Actor(),

Messages

Providing other objects with a way to use the objects.

getName(), line 14; setName(), line 19; getStageName(), line 24; setStageName(), line 29; and toString(), line 34

Operations

Telling the objects how to behave.

lines 15, 20, 25, 30, 35 and 36

Comments

Telling programmers how to use or maintain the class (ignored by the compiler).

lines starting //, e.g. lines 1 and 4

Actor,

and stageName, line 5 line 8

In addition to the elements listed above, pure object-oriented programming languages allow the programmer to specify which parts of a system can access the elements: we can usually specify, at least, that elements are public (visible everywhere) or private (only available to the objects themselves) – hence the public and private keywords in Java. Some languages allow the programmer to add assertions – logical statements that must always be true, such as ‘Objects of this class will always have a positive balance’ or ‘This message will always return a nonempty string’. Assertions are useful for reliability, debugging and maintenance.

2.16 SHARED DATA AND SHARED OPERATIONS In an object-oriented program, all the information and services the program needs must be available somewhere. If the program has been designed properly, the information and services will be available in obvious places. Bearing this in mind, where would you place the information and services listed below? 1. The current interest rate for savings accounts. 2. The number of days in January.

39

40

Chapter 2

3. The calculation of compound interest for a given number of years. 4. The calculation of whether the current year is a leap year. You may have concluded that each of them should be associated with an object of some kind – a SavingsAccount for cases 1 and 3 and a GregorianCalendar for cases 2 and 4. However, none of these pieces of information or services are related to a particular SavingsAccount or GregorianCalendar. Rather, they’re related to all savings accounts and all calendars. It would seem inappropriate to place these elements on particular objects. (If you prefer the practical argument, think of the waste of space if we had to place an interestRate on every instance of SavingsAccount, or the waste of time having to create a GregorianCalendar just so that we could ﬁnd out whether the current year is a leap year.) Because information and services such as those listed above don’t seem to ﬁt well with objects, object-oriented languages usually allow the programmer to put elements onto the class itself. So, as well as ﬁeld, message and method, we have class ﬁeld, class message and class method. Java programmers, for example, can use the keyword static to indicate that an element is associated with the whole class rather than any of its instances. Some languages go so far as to treat the class as an object in its own right. Class elements are not as easy to use as they could be, because some languages don’t treat a class as a pure object – inheritance between class elements doesn’t work, for example. Even languages that do treat classes as pure objects run into messy complications with metaclasses. So you should expect to come across class ﬁelds, class messages and class methods in other people’s designs and code and you may even ﬁnd a reason to use them yourself, but always try the following alternatives ﬁrst: • Find, or introduce, another kind of object. For example, rather than making interestRate a class ﬁeld, make it a ﬁeld on a Bank object. (This would also help you to extend your software to deal with more than one bank.) • Use a Singleton, a class that is guaranteed, by careful programming, to have only one instance: the singleton object (see Chapter 11 for more information). This is a good match for the ‘Is it a leap year?’ case, because there is only one Gregorian calendar.

2.17 TYPES In the pure object-oriented universe, everything is an object. Or, to put it in programming terms, the type of every value is a class. Smalltalk and Eiffel are two languages that stick to this fundamental rule. However, most object-oriented languages also have nonobject types, called primitives. The reasons given for this impurity usually include brevity, performance and ancestry. A cynic would suggest that laziness (on the part of the language designers)

Terminology

might also be a factor. Whichever argument you prefer, primitives are everywhere. Therefore, you need to get used to them. Taking Java as an example, we can declare that a ﬁeld has a class type such as String – a sequence of characters – or a primitive type such as int – a simple number. In Figure 2.21, the age ﬁeld of anActor is an integer primitive and the string object pointed to by name contains primitive characters. anActor age = 56

name

aString "Topol"

Figure 2.21: Object-oriented types

The main distinction between objects and primitives is that, although primitives can be used as values in the same way that objects can, you can’t send a message to a primitive, or give it a ﬁeld, or treat it as an object in any other way. (At a deeper level, most of the time objects are accessed via pointers while primitives aren’t.) You might expect that treating some things as objects and others as primitives would lead to all sorts of confusion, but it turns out not to be a major issue. Just remember that primitives are good for simple values like numbers and individual characters, but everything else should be an object. Most languages provide us with a handful of ready-made primitive types. For example, Java gives us byte, short, int, long, ﬂoat, double, char and boolean. Even if your language does allow you to deﬁne your own primitives (for example, C++ and Eiffel), you should consider this to be an advanced technique. Arrays, denoted in Java with the [] operator, sit somewhere between objects and primitives: they’re special objects that are known to the compiler and the run-time system for efﬁciency reasons. However, if you wish to be pure, you can avoid arrays altogether and use classes such as List instead. UML contains a mechanism for deﬁning language-independent primitives with names such as Integer, Real and Boolean. However, Java primitives will be used in this book as all the code fragments are in Java. Using language-dependent types is sanctioned by the UML standard, as long as you’re clear about what you’re doing.

2.18 TERMINOLOGY There are many terms available for the object concepts we’ve seen and different people use different terms to refer to the same thing. To make matters worse, some people use terms incorrectly. Figure 2.22 shows some of the terms, grouped to show how they can be used interchangeably (the underlined words are used in this book).

41

42

Chapter 2

operation

attribute

method

field

instance variable

member function

state

property data member

activation invocation call

composition

message

aggregation execution

association delegation

Figure 2.22: Object-oriented terminology

Table 2.3: Terms used in this book Term Attribute

Deﬁnition A small piece of information – such as color, height or weight – that describes one characteristic of an object.

Field

A named value inside an object.

Operation

A piece of code belonging to an object.

Method

A synonym for ‘operation’.

Message

A request sent from one object to another.

Invocation

The carrying out of an operation in response to a message.

Execution

A synonym for invocation.

Association

A direct or indirect connection between two objects.

Aggregation

A strong association implying some kind of part–whole hierarchy.

Composition

A strong aggregation, where the part is inside exactly one whole – the part may also be created and destroyed by the whole.

Interface

A set of messages understood by an object.

Protocol

An agreed way of passing messages over a network.

Behavior

A collective term for all of an object’s operations.

Reusing Code

You may also come across collective terms such as behavior (a collection of operations), interface (a collection of messages), object protocol (a synonym for interface) and data (a collection of ﬁelds). In this book, only the terms listed in Table 2.3 will be used, in the manner described. An attribute can be stored (encapsulated) by an object, but it doesn’t have to be. For example, a circle has radius and diameter attributes, but we need only store the radius because the diameter can be calculated on demand. To avoid confusion, only stored attributes will be shown in this book – if necessary, derived attributes will be implied by the addition of one or more operations, such as getDiameter. A ﬁeld is not quite the same as an attribute. Firstly, a ﬁeld represents a decision to store something; secondly, a ﬁeld can be used to store a connection to another object, as shown by navigable links on our object diagrams. When we move into design, the attributes and associations end up as ﬁelds. In the early stages of object-oriented software development, we tend to use the terms attribute and operation (because that’s UML terminology). In the later stages, when dealing with low-level design and source code, we tend to refer to ﬁelds and methods (because that’s programming terminology).

2.19 REUSING CODE The term reuse has been mentioned several times now, so let’s take a closer look at what it means and what its implications are. To put it simply, reuse refers to using code more than once, resulting in: • Faster and simpler development. • Easier maintenance (less code to harbor faults). • More robust code (every time it is reused, it is retested and, over time, more and more faults are squeezed out). There are historical and technical reasons why reuse has taken a long time to emerge. From a historical point of view, the software industry was initially preoccupied with the problem of how to program computers at all (starting with machine code), then with how to program them more efﬁciently and then with how to develop large systems in a reliable, systematic way. The potential of reusing code in different contexts became apparent as more and more systems were constructed – millions upon millions of lines of code being written every year over many decades. Surely there couldn’t be that many distinct systems that needed writing, could there? Since most software was written in software departments behind the closed doors of individual companies, using a variety of different programming languages, it’s easy to see why reuse was rare. Even when packaged software started making it out through

43

44

Chapter 2

company doors, the source code, and even the binaries, were kept a closely guarded secret. Even so, protecting a company’s commercial advantage and intellectual property doesn’t justify failing to reuse code within the company itself. From a technical point of view, reuse was difﬁcult because of an attitude in development teams and central to the development methodology that ‘We only need to solve today’s problem’. This was exacerbated by the lack of facilities provided in programming languages and methodologies to promote reuse: there was a tendency to scatter the data and operations in small pieces throughout the system, making it difﬁcult to collect related pieces into a larger chunk that would be worth reusing. Because object orientation is driven by the modeling of general concepts from an entire domain, it’s much more likely that reuse opportunities will emerge. For example, developing a payroll system would normally involve the development of an Employee class. Since the development should be driven by ‘What does the employee concept mean to this company?’ rather than ‘What employee information does the payroll system need?’, the end result is likely to be applicable to other systems developed by the same company. In addition, the modularity of objects reduces the tendency to spread attributes and operations around the system, making it easier to extract and reﬁne the Employee code. Gradually, the situation has improved, to the extent that, often, we will only need to write the parts of the system that are unique to our problem – the rest of the system can be implemented using prewritten code. This is especially true for common, well understood, application areas such as: user interfaces; database access; distributed programming; input/output; network access; e-commerce; access to legacy (pre-existing) systems; security (authentication, authorization, privacy, integrity checking, origin checking); text processing; mathematics; games; service look-up; sound synthesis and playback; 2D graphics; 3D graphics; e-mail; image processing; multimedia encoding and decoding; messaging; transaction processing; telephony; speech synthesis and recognition; and integration with digital TV broadcasts. Reuse opportunities can be summarized into the categories listed below: • Reusing functions within a system: The simplest form of reuse (used in traditional systems development) involves writing utility functions that are called from various places. For example, you may discover that various parts of your system need to search through a list of customer names, so you write a general search function that can be called from each context. Writing reusable functions is different from writing functions that break a complex process into simpler steps. • Reusing methods within an object: Methods encapsulated within an object can be invoked from other methods. For example, a nonpublic drawFilledRectangle method inside a GUIComponent class can be used by any GUIComponent method that needs to ﬁll an area of the screen with the current background color. You should aim to reuse methods within an object whenever possible. Nonpublic methods within an object are often used to break up a complicated process, in traditional fashion.

Reusing Code

• Reusing classes within a system: Many of the classes that we deﬁne can be used in different parts of our system. For example, if you deﬁne a Customer class for use in a marketing system, you expect the same Customer object to appear in many different pieces of system code. This kind of reuse is fundamental to the object-oriented approach. • Reusing functions across systems: General functions can be reused (in traditional systems development as well as in object-oriented development) in other systems that you and your colleagues produce. For example, you might write a function that extracts the year that an employee joined the company from the employee’s payroll number. For such a function to be reused by your colleagues, you would have to make them aware of it, preferably by putting it in a reuse repository: a database of useful functions that developers are expected to peruse when they’re writing new code. • Reusing classes across systems: We can publish and reuse a whole class (with all its attributes and operations) rather than just a single function. An example would be an Employee object that encapsulates the employee attributes used throughout the company, along with a useful set of operations. Object-orientation enthusiasts were the ones who popularized the idea of a reuse repository containing classes rather than functions. In order to get developers to put in the extra effort required to make their classes reusable enough for a repository, it’s a good idea to offer some kind of reward (a mug with a humorous logo on it is a common choice). Access to class repositories may be offered to third parties, perhaps for a fee. Reusing classes across your own systems is not difﬁcult, once you have a little practice. • Reusing classes across all systems: A software component is analogous to a hardware component. Software components are designed to be reusable in any context; are strongly encapsulated (clients can’t see the inner workings); come with a standard style of interface; and are available from third parties, usually in return for payment. Every object-oriented programming language has its own form of software component, for example, Java has JavaBeans. There is no real equivalent of software components in the traditional arena, because a handful of related functions would not be of much use to third parties. Examples of software components include a spreadsheet that can be dropped into any ofﬁce productivity suite and an income tax object that can be dropped into any home tax package. A software component is really just an object that obeys sensible rules about its style of interface (as in using naming conventions to identify getters and setters). • Function libraries: Related, high quality functions can be grouped into a library, so that they’re available all at once. An example would be the stdio function library, originating on Unix systems, that provides input/output facilities for C programmers. Function libraries are used in both traditional systems development and object-oriented development. Well-designed libraries sometimes become standardized by bodies such as ISO or ANSI. Function libraries may be internal to a company, free, or sold for proﬁt.

45

46

Chapter 2

• Class libraries: An improvement on function libraries, class libraries offer whole classes rather than mere functions. Writing a class library requires lots of experience. A good example is the Java 2 Enterprise Edition (J2EE) library [Bodoff et al. 02], which provides code for all of the well-understood reuse areas listed above. Just like function libraries, class libraries may be internal to a company, free, or sold for proﬁt. • Design patterns: A design pattern is a description of how to create part of an object-oriented system elegantly and effectively. Since their introduction, patterns have also been applied to other areas such as system architectures. Each pattern has a short description, a detailed description, advice on where to use it, and code samples (see Chapter 11). For example, the description of the Adapter pattern is ‘An adapter translates the interface of one object into another interface that clients expect’. Designing patterns requires a lot of experience, but less than producing a class library. • Frameworks: A framework, as its name suggests, is a pre-existing structure to which you attach your own code. In the object-oriented case, a framework consists of a number of prewritten classes, along with a document describing the construction rules that must be followed by the developer. A large example is the Enterprise Java Beans (EJB) framework [Bodoff et al. 02]: this consists of the J2EE library plus a document, hundreds of pages long, that speciﬁes how programmers should write reusable enterprise components and how third parties should implement Java application servers. Most frameworks are designed by gurus (a guru is an expert’s expert). So, how should you design for reuse? We’ll leave aside the issues of designing patterns and frameworks, since such techniques are beyond the scope of this book, and concentrate instead on writing reusable classes. Even a single reusable class will often have one or two closely collaborating classes – thus, we might produce a small family of reusable classes rather than just one. Here are some tips: • Always follow style guidelines: Style guidelines are recommendations for how you should write your classes. If you write your classes in an exotic or personal style, potential reusers of your software will quickly move on to look at other code, rather than learn your personal idiosyncrasies. Style guidelines may originate from your company, or they be more widely accepted. For example, since Sun has control over the Java standard, whatever Sun says is good style is normally accepted by the Java community. As well as having a plethora of object-orientation gurus and experts of its own, Sun pays close attention to the opinions of external experts and gurus. • Be thorough with your documentation: Few programmers will be able to understand how your classes should be reused just by reading the source code. At the very least, your class should have an explanatory name, a short comment (one or two sentences)

Reusing Code

•

•

•

•

•

•

summing up the class, a longer comment (maybe several paragraphs) describing how the class should be used and a short comment alongside each public message describing how the method should be used. Your comments should always describe the contract between the object and its clients, setting out the obligations on both sides. Documentation that is separate from the classes, such as design or tutorial information, should also be provided. Be prepared to write more code than you need: Often, when you’re implementing a class for a particular system, you will ﬁnd yourself thinking along the lines of ‘I bet a foo method would come in handy here’. For example, even if you currently only need a getRadius message for your Circle class, it would be a good idea to add a getDiameter message too, making the class more useful in other systems. Use patterns and frameworks: Patterns and frameworks reduce your workload, but they’re also understood by other developers, which means that the other developers will have less to learn before reusing your code. Design client–supplier objects: If you have two-way, or even cyclical, collaborations between your objects, you may end up with what is referred to in the trade as code spaghetti. Things will be a lot simpler if you design your objects as a client–supplier hierarchy. For example, a reusable Employee does not make any assumptions about its context; instead, it provides public messages that are generally useful. Taking this idea further, the Employee would be in control of its EmploymentHistory, but the EmploymentHistory would not know about the Employee. Imagine that your objects are servants that do what they’re told, without caring about who asked them to do it. To compensate for their servile lives, the objects have servants of their own to control. Make each object single-purpose: This is referred to as high cohesion. Avoid coding objects that serve multiple purposes, such as maintaining an employee’s personal information as well as their employment history. Separate the interface from the business behavior: A reusable object should be usable in any context. For example, the object might need to be used, directly or indirectly, in many different kinds of interface (workstation, mobile phone, or Web server). If you pollute the object with the details of a particular interface, you will run into problems. Therefore, write business objects that contain only business behavior. You can also provide interface objects to view your business objects, but that’s optional. Such interface objects will become reusable in their own right. Design for questions and commands: Objects are simpler if their messages are either questions – ‘What’s the time?’ – or commands – ‘Set the time to . . . ’ Although messages that are questions and commands at the same time are occasionally useful, you should consider their use to be an advanced technique. Combined messages, such as ‘Set the time to . . . and tell me what the time was before I set it’, can be confusing.

47

48

Chapter 2

2.20 SUMMARY In this chapter, we have looked at: • Software objects, which represent real-world things, are described by attributes and can carry out behavior (operations, usually called methods). • Messages, which enable objects to communicate and collaborate to accomplish some task. • Garbage collection, which reclaims the space used by objects when they are no longer needed by the program that created them. • Classes, which enable us to group similar objects and share the deﬁnition of elements between related objects, so that we don’t have to repeat ourselves. • Reusing code, which results in faster and simpler development, more robust code and easier maintenance.

FURTHER READING One of the original books on object concepts, and why they’re a good thing, is by David Taylor. [Taylor 97] is aimed at nontechnical readers, so it constitutes a gentle introduction for those who won’t end up producing the actual code. In order to promote the use of Java, there is a huge amount of free information on Sun’s web site at http://java.sun.com, including tutorials. Because Java is continually being updated and improved, Sun’s web site is an essential resource for Java developers. If you need help with the code fragments in this book, check out the on-line language tutorial, which is also available as [Campione et al. 00].

REVIEW QUESTIONS 1. In a UML diagram, how are objects distinguished from classes? Choose only one option. (a) Object labels are shown in italics. (b) Class labels have a box drawn around them. (c) Object labels are underlined.

Review Questions

Car color: Color

Car

Diagram 1

color 1

Color

Diagram 2

Figure 2.23: For use with Review Question 2

2. In Figure 2.23, what do Diagrams 1 and 2 illustrate? Choose only one option. (a) (b) (c) (d) (e)

1: An aggregation, 2: A composition. 1: An attribute, 2: An aggregation. 1: An aggregation, 2: An attribute. 1: An attribute, 2: A composition. 1: A composition, 2: An attribute.

3. What is meant by ‘object identity’? Choose only one option. (a) (b) (c) (d)

Two objects are identical if their attributes have the same value. Every object’s class has a unique serial number. All objects are the same as each other. Every object has a unique identity that distinguishes it from all other objects.

4. Which of the following terms best describes an object that is made up of other objects? Choose only one option. (a) (b) (c) (d) (e)

Generalization. Inheritance. Association. Aggregation. Specialization.

5. What is ‘encapsulation’? Choose only one option. (a) (b) (c) (d)

Depicting objects using doughnut diagrams. Ensuring that the data inside an object can only be accessed via operations. Sealing the state of an object so that it cannot be changed. Putting objects into a collection.

49

50

Chapter 2

ANSWERS TO ACTIVITY 1 1. A street exists independently of its houses. Although the street and the houses may be constructed at the same time, as time passes new houses will be added and old ones will be knocked down. Even if there are no houses at all, the street is still a street. This must be association. 2. If the book is a novel, we can’t tear out a page and still have a useful novel, so this is probably aggregation. If, on the other hand, it is a reporter’s notebook with pages that were designed to be torn out, it might be considered an association. 3. Notes in a symphony are similar to the pages in a novel, i.e. aggregation. If the symphony is being performed and the orchestra accidentally drops a note, people are going to notice – the orchestra members will be accused of not playing the whole symphony. 4. It may be tempting to think that components of the home entertainment system are put together to form a whole (this is implied by the use of the word ‘components’). However, if we move the games console into the kids’ bedroom, we’re still left with a home entertainment system, albeit a less entertaining one. Also, we could add a DVD player: this wouldn’t suddenly transform what we already have into a home entertainment system, it would just improve it. This looks like an association. Aggregation is still involved: the television is an aggregation, as is each of the other components. Therefore, we can have aggregations inside associations. We can also have associations inside aggregations (consider a DVD being played by a DVD player). Whether we view something as an aggregation or an association often depends on the scale we’re looking at.

ANSWERS TO REVIEW QUESTIONS 1. In a UML diagram, objects are distinguished from classes because c. Object labels are underlined. 2. In Figure 2.23, Diagrams 1 and 2 illustrate option d. 1: An attribute, 2: A composition. 3. ‘Object identity’ means that d. Every object has a unique identity that distinguishes it from all other objects. 4. An object that is made up of other objects is described as d. Aggregation. 5. ‘Encapsulation’ means b. Ensuring that the data inside an object can only be accessed via operations.

3 Inheritance

A

lthough inheritance is widely considered essential, it’s rather more complicated than what we’ve seen so far, so this chapter can be skipped on a ﬁrst reading. It is quite reasonable to develop software without the use of inheritance. However, if you do skip this chapter, you will have to accept a certain lack of understanding of some of the topics later in this book.

Learning Objectives Understand what we mean by inheritance.

Understand the difference between abstract and concrete classes.

•

Know when to use inheritance.

Chapter Outline 3.1 Introduction 3.2 Designing a Class Hierarchy 3.3 Adding Implementations to a Class Hierarchy 3.4 Abstract Classes 3.5 Redeﬁning Methods 3.6 Implementing a Stack Class 3.7 Multiple Inheritance 3.8 Guidelines for Using Inheritance 3.9 Summary Further Reading Review Questions Answers to Review Questions

3

54

Chapter 3

3.1 INTRODUCTION Inheritance allows us to specify that a class gets some of its characteristics from a parent class and then adds unique features of its own – this leads to the description of whole families of objects. Inheritance allows us to group classes into more and more general concepts, so that we can reason about larger chunks of the world that we live in. From a programming point of view, we want inheritance because: • It supports richer, more powerful, modeling. This beneﬁts both the development team and other developers who might want to reuse code. • It allows us to deﬁne information and behavior in one class and share the deﬁnitions in related subclasses. This means that we have less code to write. • It’s natural. This is one of the prime motivations for object orientation in the ﬁrst place. A subclass inherits all of the ﬁelds, messages, methods (and assertions) of its superclass. For example, if we wanted to model land vehicles, we might come up with the hierarchy shown in Figure 3.1. Car milesPerGallon:float getMilesPerGallon():float setMilesPerGallon(:float)

Bicycle

LandVehicle topSpeed:float

numberOfGears:int

getTopSpeed():float setTopSpeed(:float)

getNumberOfGears():float setNumberOfGears(:int)

Truck axleWeight:float getAxleWeight():float setAxleWeight(:float)

Figure 3.1: What is inherited?

Designing a Class Hierarchy

In this diagram, the types of ﬁelds, message parameters and message replies have been shown. In UML, types are shown after a colon, for example, :String. For brevity, parameter names have been omitted from Figure 3.1: this is only reasonable because there are few parameters and their meaning is obvious. It does, however, show the type of parameter and we must still include the colon, otherwise the label is ambiguous; so :ﬂoat indicates a ﬁeld or parameter of type ﬂoat, whereas ﬂoat indicates a ﬁeld or parameter with the name ﬂoat. If parameter names had been included, they would have looked something like newTopSpeed:ﬂoat. Using this class hierarchy, if we were to create a Car, it would have one ﬁeld inherited from LandVehicle – topSpeed – and another ﬁeld introduced by the Car class itself – milesPerGallon. It would also have four methods (getTopSpeed, setTopSpeed, getMilesPerGallon and setMilesPerGallon), two inherited and two introduced by Car. If we were to create a Bicycle, it would have the TopSpeed elements, just the same as a Car, but it would have the numberOfGears elements instead of the milesPerGallon elements. So, the hierarchy shown, as well as allowing us to reason about general classes and speciﬁc classes, reduces our programming effort (because all the TopSpeed elements from LandVehicle, for example, appear automatically in its subclasses, without us having to repeat them).

3.2 DESIGNING A CLASS HIERARCHY Let’s look at a larger example. We want to model collections, objects that can hold on to other objects for later use. After some deliberation, we decide that we need four styles of collection: • List: A collection that keeps all of its objects in the order in which they were inserted. • Bag: A collection that doesn’t keep its objects in order. • LinkedList: A collection that keeps its objects in order using an implementation of a sequence of objects in which each object points to the next in the sequence. A linked list can be updated easily, but access is slow because we have to walk down the list. • ArrayList: A collection that keeps its objects in order using an array, a sequence of adjacent memory locations. Arrays have fast access but updating is slow because we may have to shift elements around or create a new array on each update. How could we place these four classes, along with Collection, into an inheritance hierarchy? The key is to look for major similarities between the concepts. Clearly, they’re all collections in their own right, so Collection must go at the top. Next, we notice that most of the collections keep their objects in order, but Bag doesn’t – this suggests that Bag should be placed directly under Collection, in a separate branch from the other classes. Next, we notice

55

56

Chapter 3

Collection

Bag

List

ArrayList

LinkedList

Figure 3.2: A collection hierarchy

that List has no commitment to internal implementation, whereas LinkedList and ArrayList do (the only difference between LinkedList and ArrayList is their time–space trade-off). So, we decide that List must be a superclass of ArrayList and LinkedList. This logical process leads to the hierarchy shown in Figure 3.2. The inheritance arrows have been drawn a little differently to those in the previous chapter; this form, also an ofﬁcial part of UML, is neater and more compact. The process just described is somewhat artiﬁcial, because we had all the classes in the hierarchy to begin with, and then built the result from the top down. In reality, it tends to happen the other way round: ﬁrst, we decide what classes we want at the bottom of the hierarchy – ArrayList, LinkedList and Bag – then we look for more general concepts, so that we can enrich our model and share element deﬁnitions. Thus, we might group ArrayList and LinkedList into List and then group Bag and List into Collection. While we’re developing a hierarchy, we look for messages that we can share – the higher up the hierarchy we can place the messages, the better. We tend to look for messages before we look for any other class elements because the messages represent the interface that our objects will present to the outside world, their most important feature. Now consider the following three messages, which we’ve decided are good candidates for placement in our collection hierarchy: • contains(:Object):boolean searches for an object in the collection and returns true if the receiver contains the parameter or false otherwise.

Designing a Class Hierarchy

• elementAt(:int):Object retrieves the object at the position indicated by the parameter. • numberOfElements():int replies with the number of objects in the collection. Where can we place these messages on the classes that we already have? Well, contains is something that we should be able to ask any kind of collection, so we need to guarantee that by putting it on Collection. Now, elementAt takes a position as parameter, so it must be dealing with ordered objects (it wouldn’t make sense for Bag). So, we can’t put elementAt on Collection; we could put it on ArrayList and repeat it on LinkedList, because both of these classes keep their objects in order. However, we can avoid the repetition by putting it on List instead. Finally, numberOfElements is something we might want to ask of any collection, so the Collection class is the appropriate place to put it. These deliberations lead us to the distribution of messages shown in Figure 3.3. Even though attributes have not been shown in this class diagram, the attribute box has been retained – this makes it easier to see that the list shows messages rather than ﬁelds. The message names are in italics to emphasize that we are not yet considering methods.

Collection contains(:Object):boolean numberOfElements():int

Bag

List elementAt(:int):Object

ArrayList

Figure 3.3: Placing messages in a hierarchy

LinkedList

57

58

Chapter 3

We now have a complete set of classes that would be valuable to other developers. However, the classes don’t yet do anything: they have no implementation.

3.3 ADDING IMPLEMENTATIONS TO A CLASS HIERARCHY Okay, so we have our class hierarchy and we’ve decided where the messages should be introduced. Now the implementation elements (ﬁelds, constructors and methods) must be added. At this stage, let’s not worry about the ﬁelds that our hierarchy would need, because that would be getting into detailed design. Similarly, let’s ignore the issue of constructors for our classes – we can just assume that our ﬁnished classes will be furnished with a useful and complete set of initialization choices. Instead, we’ll look at the issue of where to put the methods, because that will lead us on to two important concepts: abstraction and redeﬁnition. We conclude that it’s impossible to write a contains method on the Collection class because the search algorithm is going to be different for ordered and unordered collections. So we must, at least, implement a contains method on Bag. But how about List and its subclasses? With a little thought, using the other messages that we’ve introduced, we can write a contains algorithm that works for any kind of List (see Implementation Point 3).

Implementation Point 3 1 2 3 4 5 6 7 8

boolean contains(Object o) { for (int i = 0; i < numberOfElements(); ++i) { if (elementAt(i) == o) { return true; } } return false; }

Notice that parameters in Java have their type followed by their name (see line 1), unlike parameters in UML which are given as name:type. The for-loop (line 2) sets i to every value from 0 up to, but not including, the number of elements in the list. Inside the loop, the current object is retrieved and tested (line 3). Writing aMessage without specifying the object that is to receive the message means ‘send aMessage to the current object’ (some programmers prefer to write this.aMessage() instead). For example, elementAt(i) means ‘ﬁnd the value of the ith element of the current List object’.

Abstract Classes

Now we’re really reaping the beneﬁts of inheritance: we’ve written a single method that will work perfectly for any direct or indirect subclass of List, which could be many classes. The elementAt message is going to have a different implementation for ArrayList and LinkedList. So, we must add two separate elementAt methods: one to ArrayList, which will access the elements directly; and one to LinkedList, which will walk down the list. Last, but not least, we have to place numberOfElements. This implementation depends on whether we store the value as a ﬁeld or calculate it when asked. Let’s look at each alternative: • Storing the number of elements as a ﬁeld The ﬁeld has to be incremented whenever we add an object and decremented whenever we remove an object. This approach allows us to report the number of objects quickly, at the cost of extra storage and slightly slower operations to add and remove objects. • Calculating the number of elements on demand This is likely to be very slow for a LinkedList, involving walking down all the elements. For ArrayList and Bag, the internal objects would probably store the number of elements anyway, so the operation would be fast. Either way, with this approach, we don’t waste storage and we don’t slow down the adding and removing of objects. There is a time–space trade-off that the designer needs to resolve. On this occasion, for example, we decide that neither option works well for all classes, nor even for all List objects. Therefore, we decide to put a separate numberOfElements method on each of the three leaf classes, Bag, ArrayList and LinkedList. Having made our implementation decisions, we arrive at the hierarchy shown in Figure 3.4. So that you can distinguish between them, we’ve shown messages in italics and methods in roman text (which is also correct UML notation). On our Collection classes, we now have two styles of message: those with an associated method and those without. A message without a method is unﬁnished – we know that the associated class has the message as part of its interface, but we have to look down the subclass chain to ﬁnd the actual method. An unﬁnished method is called an abstract method (or abstract operation), because it’s not real, it’s not solid, you can’t kick it. The complement to an abstract method is a concrete method (or concrete operation). A concrete method has real lines of code, it’s solid, you can kick it. In UML, abstract methods are shown in italics and concrete methods are not – where italics are impractical, you can put {abstract} to the right of the method instead.

3.4 ABSTRACT CLASSES An abstract class is a class with at least one abstract method – the abstract method may be introduced on the class itself, or it may be inherited from a superclass. Figure 3.5 shows

59

60

Chapter 3

Collection contains(:Object):boolean numberOfElements():int

Bag

List

contains(:Object):boolean numberOfElements():int

contains(:Object):boolean elementAt(:int):Object

ArrayList

LinkedList

elementAt(:int):Object numberOfElements():int

elementAt(:int):Object numberOfElements():int

Figure 3.4: Placing methods in a hierarchy Collection contains(:Object):boolean numberOfElements():int

Bag

List

contains(:Object):boolean numberOfElements():int

contains(:Object):boolean elementAt(:int):Object

ArrayList elementAt(:int):Object numberOfElements():int

LinkedList elementAt(:int):Object numberOfElements():int

Figure 3.5: Abstract class names in italics

Abstract Classes

our collection hierarchy again, with the names of the abstract classes in italics (this is also correct UML). Again, if italics are impractical, you can write {abstract} above or to the left of the class name. Abstract classes have the following advantages: • They permit richer and more ﬂexible modeling; for example, our List class has all three messages – contains, elementAt and numberOfElements – despite the fact that we can’t provide concrete methods for all of them. • They lead to more code sharing, because we can write concrete methods that use abstract methods; for example, the contains method for List invokes abstract methods. Once again, abstraction is a natural thing to want to do, so object orientation provides the facility. Consider peeling a piece of fruit: we know that we can take the skin off any piece of fruit, but we can’t describe how to do it in a way that will work well for every variety. Therefore, peeling fruit must be an abstract method and fruit itself must be an abstract concept. With the abstraction facilities presented so far, we have a disaster waiting to happen: what do you think would happen if we created a Fruit and then sent it the peel message? How about if we created a Collection and sent it the contains message? We would be trying to invoke an abstract method – a thing that doesn’t really exist – so the objects wouldn’t know what to do. Most object-oriented languages stop us creating instances of abstract classes. For example, a Java compiler won’t compile any program containing the expression new List(). The compiler is being paranoid in this case, because we might never invoke an abstract method on the new object, but the compiler can’t check that for us. The rule that instances of abstract classes can’t be created corresponds well to the real world: if I give you some money and ask you to go to the shop to buy a piece of fruit, you’re likely to ask ‘What kind of fruit would you like?’, because you need a concrete request. When designing a class hierarchy, you should bear in mind that most superclasses are abstract. This follows from the fact that inheritance hierarchies are naturally derived from the bottom up: 1. We look for the concrete concepts that exist in our problem domain and reason about their knowledge and behavior. 2. We look for commonalities between the concrete classes so that we can introduce more general superclasses. 3. We group superclasses into more superclasses, until we arrive at our most general root class (Fruit or Collection, for example).

61

62

Chapter 3

X

X1

usually ends up as Y

Y

X2

Figure 3.6: Most superclasses are abstract

When we identify generalizations (superclasses), we expect them to be abstract – otherwise it’s likely that they would have been identiﬁed as concrete concepts in Step 1. Most of the time, the observation that most superclasses are abstract is borne out by the discovery, sooner or later, of abstract methods on the superclasses (possibly inherited). In order to help with this, Java and UML allow us to mark a class as abstract even if we haven’t yet discovered any abstract methods for it. So, what do you do if you ﬁnd yourself with a concrete class inheriting from another concrete class? One thing you can try is the transformation shown in Figure 3.6. Here, we have a concrete class called Y inheriting from another concrete class X; we can transform this into an abstract superclass X1, with two concrete subclasses, X2 and Y. If you try to work with the hierarchy on the left, you will tend to ﬁnd that Y is not a true subclass of X – it doesn’t quite work as a kind of X, leading you to change the meaning of some methods, or even to disable them. Imagine if X were Fruit and Y were Orange; we would expect an Orange to work exactly like a Fruit, with extra knowledge and behavior added on, so the disabling of methods would be unacceptable. For the hierarchy on the right, every piece of knowledge and behavior that is truly common to X and Y has been put in X1; everything in X that is not common to Y has been moved into X2. (The old X has been split in two.) We now have a much cleaner picture: X1 is likely to be abstract; X2 and Y behave exactly as kinds of X1; X2 and Y have their own extra knowledge and behavior. Sometimes, we’re tempted to inherit from a concrete class, in order to tweak the elements of something that already exists: for example, we may already have an ArrayList class implemented by another developer, but it doesn’t quite do what we want it to do, so we introduce a new subclass called MyArrayList and add minor modiﬁcations. Although this is sometimes reasonable, it has one fundamental drawback: any existing code that creates instances of ArrayList will still do so: there’s no way that we can force the existing code to create instances of our improved class.

Redeﬁning Methods

3.5 REDEFINING METHODS Object orientation allows us to redeﬁne elements that we inherit. In its simplest form, redeﬁnition allows a subclass to change the implementation of an inherited method – the message stays the same but the lines of code are replaced. Another form of redeﬁnition allows us to make a message more visible in a subclass: since we’ve only seen public and private visibilities up to this point, this means that a subclass can turn a private message into a public message. Yet another form of redeﬁnition allows us to change the name or type of an attribute. For the rest of this discussion, let’s concentrate on redeﬁning the content of methods, since that’s the most important reason for redeﬁnition. There are three good reasons why we would redeﬁne a method: • The inherited method was abstract and we want to make it concrete, by giving it some code – for example, contains is abstract on Collection but we need it to be concrete on Bag and List. • The method needs to do some additional work in the subclass – for example, a toString method would have to summarize any new attributes that were introduced by the subclass. • We can provide a better (more efﬁcient or more accurate) implementation for the subclass – for example, if we add an index to our LinkedList class, we can redeﬁne contains to be faster than the linear algorithm used by List. When we are just doing additional work, we should make sure that the superclass deﬁnition still does everything it used to – this increases code sharing and simpliﬁes maintenance (for example, if we modify the superclass deﬁnition, the subclass gets the new behavior automatically). Every object-oriented language allows a redeﬁned method to invoke the one on its superclass (see Implementation Point 4). Implementation Point 4 Here’s a Java example: void initialize() { // Invoke the inherited initialize method super.initialize(); // Now do the extra stuff ... }

When you redeﬁne an element, do not change its meaning. When implementing a subclass, you mustn’t forget your contract with users of the superclass: anyone has the right

63

64

Chapter 3

to assume that your subclass works exactly like the superclass, with additional knowledge and behavior. For example, when redeﬁning contains for LinkedList, it wouldn’t be fair to make it return true when the parameter is not in the list.

3.6 IMPLEMENTING A STACK CLASS Let’s look more closely at the idea of reusing code by sharing. With a Stack (see Figure 3.7), we can push an object onto the top, peek at the top object, see if the stack is empty, and pop an object off the top. So, what we need is a Stack class with the following four messages: • • • •

to add an object to the top of the stack. to return the object on the top of the stack. isEmpty():boolean to return true if there are no objects on the stack. pop():Object to remove an object from the top of the stack and return it. push()

peek():Object

Figure 3.7: A spring-loaded plate dispenser

Implementing a Stack Class

Now, object-oriented software development is supposed to be about not writing code, so the ﬁrst thing we should do is to look around for a similar class that already exists. During our search, we come across the LinkedList class, part of which we developed earlier. We discover that it has four messages that we ought to be able to reuse: • • • •

which adds an object to the end of the list. lastElement():Object which returns the object at the end of the list. numberOfElements():int which returns the number of objects in the list. removeLastElement() which removes the object at the end of the list. addElement()

If we view a stack as running from left to right, instead of bottom to top, it looks as though all the methods we need are already written, but with the wrong messages. We must decide how to incorporate the existing LinkedList behavior into our new Stack class. We could do it by inheritance (see Figure 3.8) or by composition (see Figure 3.9). Composition is a strong aggregation where the composed object is inside a single composite; the composed object is usually created at the same time as the composite and can be deleted at the same time. In UML, in order to emphasize that composition is stronger than aggregation, we use a black diamond instead of a white one.

3.6.1 Implementing a Stack using Inheritance Let’s say that we choose to make our new class a subclass of LinkedList, as shown in Figure 3.8. Next, we deﬁne our stack messages in terms of the inherited messages. The following code shows four such deﬁnitions (in Java, extends means ‘inherits from’): public class Stack extends LinkedList { public void push(Object o) { addElement(o); } public Object peek() { return lastElement(); } public boolean isEmpty() { return numberOfElements() == 0; } public Object pop() { Object o = lastElement(); removeLastElement(); return o; } }

65

66

Chapter 3

LinkedList

Stack

Figure 3.8: Stack by inheritance

We can now create a stack object and use it: Person aPerson = new Person(); Stack aStack = new Stack(); aStack.push(new Plate("Wedgwood")); aStack.push(new Plate("Royal Doulton")); aStack.push(new Plate("Domestic green")); aPerson.take(aStack.pop());

However, we have a potentially serious problem with this way of implementing Stack. Since Stack is a sublass of LinkedList, all other LinkedList messages are also available to stack objects. That should be no surprise, since a subclass is expected to offer at least the same services as its superclass. Our problem is that LinkedList has messages, such as ﬁrstElement, that are inappropriate for stacks, but a client programmer would still be able to use them, for example: aPerson.take(aStack.firstElement());

This piece of code means that the client of aStack has just removed the element from the bottom of the stack – something that they’re not supposed to be able to do.

3.6.2 Implementing a Stack using Composition Figure 3.9 shows a Stack implemented with an internal reference to a LinkedList. The behavior of the encapsulated LinkedList is used by Stack, but none of the extra LinkedList behavior is exposed to Stack clients. Another way of describing this situation is to say that aStack delegates its behavior to aLinkedList. Since aLinkedList is encapsulated, the only reference to it is inside aStack. Therefore, if we’re using a pure object-oriented language with a garbage collector, aLinkedList is deleted as soon as aStack is deleted.

Implementing a Stack Class

Stack

LinkedList

push(:Object) pop():Object isEmpty():boolean peek():Object

list

add(:Object) lastElement():Object removeLastElement() numberOfElements():int

list aStack

aLinkedList

Figure 3.9: Stack by composition

Using composition, Stack can be implemented like this: public class Stack { private LinkedList list; public Stack() { list = new LinkedList(); } public void push(Object o) { list.addElement(o); } public Object peek() { return list.lastElement(); } public boolean isEmpty() { return list.numberOfElements() == 0; } public Object pop() { Object o = list.lastElement(); list.removeLastElement(); return o; } }

This class provides all the messages that a stack needs, but it does not work with: aPerson.take(aStack.firstElement());

because there is no longer a ﬁrstElement message.

67

68

Chapter 3

To achieve this improved behavior only meant that the Stack class had to declare and create a ﬁeld and the methods had to reference that ﬁeld. Although there may be a slight overhead in creating the delegated ﬁeld and forwarding messages to it, it shouldn’t be enough to concern us, especially considering the sophistication of today’s compilers and run-time systems.

3.6.3 Inheritance versus Composition Inheritance has some unique advantages: • It’s natural. • It’s elegant. • It allows us to write generic code – for example, code written to work for Fruit will also work for Apple and Pear. However, inheritance suffers from the following problems: • • • •

It’s difﬁcult to do well. It’s difﬁcult to change when you discover deﬁciencies in your design. It’s more difﬁcult for client programmers to understand. The hierarchy ‘leaks’ into client code, making it more difﬁcult to change too.

Composition achieves the same end result as inheritance (concrete classes, concrete messages and reuse of existing code). However, it has the following advantages: • • • •

It’s simpler to produce. It’s easier to change. It’s easier for clients to understand. It doesn’t leak into client code.

On the whole, especially for beginners, composition wins over inheritance: it’s reasonable to implement a sizeable application with no inheritance at all. Inheritance is best left to the experts, especially for when they’re implementing large libraries of reusable code (even then, inheritance works best for well-understood domains such as graphical user interfaces, databases, networks and collections). Some languages (such as Eiffel and C++) permit private inheritance, also called implementation inheritance. Private inheritance, as its name suggests, allows one class to inherit from another without the inherited elements becoming part of the new interface. For example, with such a facility, our Stack class could inherit privately from LinkedList: the Stack methods

Multiple Inheritance

would have direct access to LinkedList methods, without having to introduce a delegate, so there would be no arguments about convenience or efﬁciency; but, because the inheritance would be private, we would not be allowing clients of Stack to use inappropriate LinkedList messages, such as removeFirstElement, so the purists would be happy too. Private inheritance is not available in all languages, so it’s another one of those features that you should avoid during analysis and design, if you don’t want to tie yourself to a particular language. Private inheritance is usually provided as a side effect of multiple inheritance, which is also not supported by many languages.

3.7 MULTIPLE INHERITANCE When we design an inheritance hierarchy, we generalize classes into higher-level abstractions (when moving up the hierarchy) and specialize them into lower-level abstractions or concrete classes (when moving down). As we move up or down, we often have to choose between alternative generalizations and specializations, even though they may all seem equally valid. For example, consider the two inheritance hierarchies in Figure 3.10. In the ﬁrst hierarchy, Vehicles have been classiﬁed as Powered or Unpowered; in the second, they’ve been classiﬁed as Land, Air or Sea. So, we have alternative hierarchies, but which do we choose? The answer may depend on what your problem domain is (engine maker, globe-trotter or luggage manufacturer). Or, you may not be able to come up with an answer at all until you’ve tried out one of the hierarchies and found whether or not it is adequate. This kind of dilemma is a side effect of single inheritance. With single inheritance, a class is only allowed to have one parent. Single inheritance works well, which is why languages like Smalltalk and Java have nothing else. But multiple inheritance, where each class has any number of parents, is also possible. (Java does have a form of multiple inheritance, but only for interfaces, explained in Chapter 8.) With multiple inheritance, we might be tempted to combine the hierarchies in Figure 3.10 into the hierarchy shown in Figure 3.11, in the hope of getting the best of both worlds. Looking at Figure 3.11, you might think that multiple inheritance is rather complicated and not worth the bother. But there are many points for and against. Advantages of multiple inheritance are that: • • • •

It is powerful. It permits private inheritance. It is closer to the real world. It allows mix-in inheritance.

69

70

Chapter 3

Train Truck Powered

JetPlane Car Ferry

Vehicle

Glider Unpowered

RowingBoat Bicycle Yacht

Train Truck Land Bicycle Car Vehicle

Air

JetPlane Glider RowingBoat

Sea

Ferry Yacht

Figure 3.10: Alternative inheritance hierarchies

Mix-in inheritance is a design style in which we maintain a single inheritance backbone for our principal classes, while permitting individual classes to inherit from one or more decorating classes, each of which adds a few simple elements. For example, Lorry may get most of its important elements from LandVehicle but may inherit an engine attribute from a simple Powered class. Disadvantages of multiple inheritance are that: • It introduces complexity (for the designer and the client programmer). • It causes name clashes.

Multiple Inheritance

Train

Land

Truck Air

Bicycle Car

Sea

Vehicle

JetPlane Glider

Powered

RowingBoat Ferry

Unpowered Yacht

Figure 3.11: Multiple inheritance – best of both worlds?

HouseProduct

EuroProduct

isPolish():boolean

isPolish():boolean

WarsawPolish isPolish():boolean

?

Figure 3.12: Name clashes

• It causes repeated inheritance. • It makes compilers more difﬁcult to write. • It makes (fast) run-time systems more difﬁcult to write.

Name clashes occur when elements with the same name, but different implementations, are inherited via different routes. For an example, consider the hierarchy in Figure 3.12. Here we have two superclasses: HouseProduct is something bought for the house; EuroProduct is something manufactured in Europe. On HouseProduct, we have a message isPolish that

71

72

Chapter 3

returns true if the associated product is used for polishing; in contrast, on EuroProduct, we have an isPolish message that returns true if the associated product is manufactured in Poland. Once we introduce the mutual subclass WarsawPolish, denoting polish manufactured in Warsaw, we have a problem: we would like the new class to have two messages with different meanings; however, they can’t both use the same name, because that would make aWarsawPolish.isPolish() ambiguous. If we rename the two messages, clients of WarsawPolish will be surprised to learn that it has no isPolish message at all, despite what it says on HouseProduct and EuroProduct.

Employee

UKEmployee

USEmployee

Assignee

Figure 3.13: Repeated inheritance

Repeated inheritance means inheriting the same element from more than one route. For example, if a getName method is deﬁned on the Employee class in Figure 3.13, Assignee inherits getName from both UKEmployee and USEmployee. In this simple case, it should be easy for a compiler to work out that Assignee should end up with one getName method. However, things get trickier if UKEmployee or USEmployee choose to redeﬁne the getName method; Assignee could end up with two getName methods (in which case, how do we deal with the name clash?) or just one (in which case, how do we choose which one to have?). The problems of writing efﬁcient compilers and run-time systems to handle multiple inheritance are difﬁcult to overcome. However, any programmer that yearns for the power of multiple inheritance is unlikely to be satisﬁed if these are the reasons given for dropping the facility altogether. In support of multiple inheritance, it should be noted that all the problems can be overcome with the help of clever designers and good programming style. However, since it may not be available in your implementation language, you may well choose to avoid multiple inheritance altogether, especially for analysis purposes. If you do use multiple inheritance during design, restrict yourself to a single-inheritance backbone, with the addition of simple mix-in classes when appropriate.

Summary

For the record, here’s a summary of the multiple inheritance facilities available in common languages: • Eiffel provides straightforward multiple inheritance, plus private inheritance and mix-in inheritance. Eiffel has a rich set of facilities for dealing with name clashes and repeated inheritance. • Smalltalk provides only single inheritance. Multiple inheritance facilities have been tried with Smalltalk, with limited success. • C++ provides some multiple inheritance facilities. However, the facilities are incomplete, poorly designed and poorly implemented, so they should be avoided. • Java provides single inheritance for classes but multiple inheritance for interfaces (abstract classes that have no methods). This is a good compromise that allows some degree of multiple inheritance modeling and mix-in inheritance (for messages only). Java has no special facilities to deal with name clashes or repeated inheritance, but a simple form of repeated inheritance is allowed.

3.8 GUIDELINES FOR USING INHERITANCE • Don’t overdo it: Don’t think that you have to use inheritance a lot, or even at all. Remember that there are alternatives, such as composition and the use of attributes (for example, a Car class with a color attribute is probably better than three classes called Car, RedCar and BlueCar). • A class should be ‘a kind of ’ its superclass(es): Whenever you subclass X to produce Y, ask yourself ‘Is Y a kind of X?’ For example, Orange is a kind of Fruit and Truck is a kind of LandVehicle, so they are valid; conversely, Potato is not a kind of Fruit and Airplane is not a kind of LandVehicle, so these would not make good subclasses. (Some developers use the terms subtyping to mean ‘I am following the guideline’ and subclassing to mean ‘I might not be’.) • A class should be an extension of its superclass(es): In a subclass, make sure that you only add new features; don’t be tempted to break the superclass contract by deleting, disabling or reinterpreting features.

3.9 SUMMARY In this chapter, we examined inheritance, a tool for code sharing and higher-level modeling: • Classes can be grouped into more general concepts and a class can get (inherit) some of its characteristics from a parent class.

73

74

Chapter 3

• An abstract class has at least one method without code (an abstract method); in a concrete class, all the methods contain lines of code.

FURTHER READING One of the best-known books on object-oriented design and programming theory is [Meyer 97], covering everything you should and shouldn’t do with objects. Meyer illustrates his ideas using his own language, Eiffel, but Eiffel’s syntax is simple and based entirely on object-oriented concepts, so it makes a good complement to the main text. Apart, perhaps, from the choice of programming language, even the most stringent theoretician would ﬁnd it difﬁcult to disagree with anything Meyer has to say.

REVIEW QUESTIONS 1. Why is the ability to redeﬁne a method important in object-oriented programming? Choose all options that apply. (a) Because it allows us to add extra work to a method. (b) Because it allows us to introduce abstract methods that are redeﬁned as concrete methods. (c) Because it allows us to provide a more accurate or faster deﬁnition in a subclass. (d) Because it allows us to disable a method in a subclass. (e) Because it allows us to change the meaning of a method. 2. Which of the following statements about multiple inheritance are true? Choose all options that apply. (a) (b) (c) (d)

It offers more modeling choices. It makes it more difﬁcult to write compilers and (fast) runtime systems. It simpliﬁes inheritance hierarchies. It solves the problem of repeated inheritance.

3. Which of the following statements are true? Choose all options that apply. (a) (b) (c) (d)

Most superclasses are abstract. Inheritance is preferable to composition. Most superclasses are concrete. Composition is preferable to inheritance.

4. In UML diagrams, how are abstract classes distinguished from concrete classes? Choose only one option.

Answers to Review Questions

(a) (b) (c) (d)

Concrete classes are shown as boxes with dashed outlines. Labels on abstract classes are shown in italics. Labels on concrete classes are shown in italics. Abstract classes are shown as boxes with dashed outlines.

5. What is an ‘abstract’ class? Choose only one option. (a) (b) (c) (d) (e)

An object. A class with no methods. A class with no concrete subclasses. A class with at least one undeﬁned message. An interface.

6. Which of the following terms best describes the case where a Stack class is implemented using an internal instance of List? Choose only one option. (a) (b) (c) (d) (e)

Association. Specialization. Genericity. Composition. Singularity.

7. With reference to Figure 3.14, what is the difference between the two diagrams? Choose only one option. (a) (b) (c) (d)

In Diagram 1, color is public but in Diagram 2 color is private. Diagram 2 indicates that the car’s color can be removed and replaced. Diagram 1 shows an abstract class and Diagram 2 shows a concrete class. None, they mean the same thing.

Car color: Color Diagram 1

color 1

Car

Color

Diagram 2

Figure 3.14: For use with Review Question 7

ANSWERS TO REVIEW QUESTIONS 1. The ability to redeﬁne a method is important in object-oriented programming: (a) Because it allows us to add extra work to a method.

75

76

Chapter 3

(b) Because it allows us to introduce abstract methods that are redeﬁned as concrete methods. (c) Because it allows us to provide a more accurate or faster deﬁnition in a subclass. 2. Multiple inheritance a. offers more modeling choices and b. makes it more difﬁcult to write compilers and (fast) runtime systems. 3. The following statements are true a. Most superclasses are abstract b. Composition is preferable to inheritance. 4. In UML diagrams, abstract classes are distinguished from concrete classes because b. Labels on abstract classes are shown in italics. 5. An ‘abstract’ class is d. A class with at least one undeﬁned message. 6. The term that best describes the case where a Stack class is implemented using an internal instance of List is d. Composition. 7. The difference between the two diagrams is d. None, they mean the same thing.

4 Type Systems

A

s with inheritance, type systems are widely considered essential but are rather complicated, so this chapter could be skipped on a ﬁrst reading. You can get away without a detailed knowledge of type systems. However, if you do skip this chapter, you will have to accept a certain lack of understanding of some of the topics later in this book.

Learning Objectives Understand what we mean by a type system.

Understand implicit and explicit casting.

Understand polymorphism.

Recognize that Java templates provide genericity.

•

•

Chapter Outline 4.1 Introduction 4.2 Dynamic and Static Type Systems 4.3 Polymorphism 4.4 Dynamic Binding 4.5 Polymorphism Guideline 4.6 Type Casting 4.7 Explicit Casts 4.8 Genericity with Templates 4.9 Summary Further Reading Review Questions Answers to Activity 2 Answers to Activity 3 Answers to Review Questions

4

80

Chapter 4

4.1 INTRODUCTION A type system is a simple concept: it’s a set of rules that stop us misusing values (primitives and objects). Usually, this is done by forcing us to declare how we intend to use a value before we actually use it – this allows compilers and run-time systems to spot potential abuses before they happen. A simple example of a type system in use is declaring that a variable will always hold a value of a particular type: int i; Employee fred;

But why would we misuse a value in the ﬁrst place? The most common reasons are: • Not understanding how the value is supposed to be used. • Because we have spelled something incorrectly. As well as preventing the misuse of values, a type system can have a couple of other beneﬁts: it can ensure that we provide some documentation of the code (‘fred is an Employee’) and it can improve run-time performance, because the compiler and run-time system have more information about what the code intends to do (compiler) or what it actually is doing (run-time system).

4.2 DYNAMIC AND STATIC TYPE SYSTEMS Type systems can either be static (‘done by the compiler’) or dynamic (‘done by the run-time system’). Both varieties of type system ensure that values are not misused by the programmer: a static type system spots the abuse at compile time while a dynamic type system waits to see if the abuse actually happens and then stops it. Smalltalk is the archetypal dynamically-typed language. In Smalltalk, a programmer often declares the expected type of a value using a naming convention. For example, to ensure that a parameter for the addEmployee method always pointed to an Employee object, the programmer might give it the name anEmployee. Anyone using addEmployee would assume from the parameter name that they were supposed to pass in an Employee. (Similar naming conventions can be applied to ﬁelds and local variables.) But Smalltalk is dynamically typed, so the client programmer still has to make sure that they don’t invoke addEmployee with, say, a Banana. Let’s assume that, inside the addEmployee method, the getPayrollNumber message is sent to anEmployee. Therefore, if we pass in a Banana, getPayrollNumber will be sent to the Banana, which is clearly nonsense. Smalltalk’s

Dynamic and Static Type Systems

reaction to this misuse of a Banana would be to generate a run-time error with the message Banana doesn’t understand payrollNumber. Smalltalk’s approach still constitutes a type system, albeit a rather simple one, because the run-time system has prevented the client programmer from abusing the Banana – the object hasn’t been forced to answer a question that it doesn’t understand. However, we have to wait until the code is run before we discover the error. Worse still, if our test code only sends messages to anEmployee that are also supported by Banana, we won’t discover the error until we write more code or run the system in a live setting. If we write the addEmployee method in Java, we are forced to declare a type for the parameter: public void addEmployee(Employee anEmployee) { ... pay = anEmployee.getPayrollNumber() ... }

Here, the programmer is telling the compiler that the parameter passed in by the client must be an Employee (and that the message doesn’t return anything, using the void keyword). Therefore, the compiler will refuse to compile the following code fragment: aPayroll.addEmployee(new Banana());

So, in Java, the abuse of Banana is stopped by the compiler, rather than by the run-time system. A Smalltalk programmer would argue that dynamic typing is the better option, because: • It makes compilation quick and simple. • The programmer can work quickly – ideas ﬂow easily from the programmer’s head into the program, because they don’t have to keep stopping to think what kind of object they must use in this context. • The lack of a static compiler encourages thorough testing. • Object-oriented code is continually being reused, so faults are always found eventually. A Java programmer would argue that static typing is a good thing because: • Compilation is still pretty quick and, anyway, programmers don’t care how difﬁcult it is to write a compiler. • It can be used to improve run-time performance. • It can be used to pick up spelling mistakes. • It forces some documentation of the code. In object-oriented terms, a static type system guarantees that it’s impossible to send a message to an object unless that object has a corresponding method.

81

82

Chapter 4

4.3 POLYMORPHISM Polymorphism is derived from the Greek word stems poly, meaning many, and morph, meaning shape. Therefore, polymorphic means ‘having many shapes’. We can apply the term separately to variables and to messages: a polymorphic variable refers to different types of value at different times; a polymorphic message has more than one method associated with it. In a pure object-oriented language, all nonprimitive variables are polymorphic and all messages are polymorphic. We’ll take a look at each in turn.

4.3.1 Polymorphic Variables The following Java declaration states that t will always point at an object of type Truck: Truck t;

Therefore, the following assignment would be valid and would lead to the situation shown in Figure 4.1: t = new Truck();

Truck t

aTruck

Figure 4.1: Attaching a Truck variable to a Truck object

Now consider the class hierarchy shown in Figure 4.2, which tells us that a Truck is a kind of LandVehicle. Given this hierarchy, we would expect to be able to treat a Truck just like a LandVehicle. Therefore, the following statement would also be valid and would give us the situation shown in Figure 4.3: LandVehicle lv = new Truck();

That may surprise you. But, consider this: we’ve just told the compiler that we will only ever use lv as a LandVehicle, i.e. we will only ever send it LandVehicle messages; since all the LandVehicle messages are on Truck too, everything will be ﬁne. In the same way that we can make lv refer to a Truck, we can make it refer to a Train (see Figure 4.4): lv = new Train();

Polymorphism

Train

Vehicle

LandVehicle

Truck

Figure 4.2: Truck inheritance

LandVehicle lv

aTruck

Figure 4.3: Attaching a LandVehicle variable to a Truck object

LandVehicle lv

aTrain

Figure 4.4: Attaching a LandVehicle variable to a Train object

Although lv was pointing at a Truck, it’s now pointing at a Train, so it must be a polymorphic variable. This should be intuitive: lv is a LandVehicle, so it can point at any kind of LandVehicle. The polymorphism of variables is controlled by inheritance. For example, because Orange is not a kind of LandVehicle, it wouldn’t be sensible to write lv = new Orange();

So, polymorphism allows us to attach a variable to a subclass object. But we can’t go the other way round. For example, because a Vehicle is not a kind of LandVehicle, it wouldn’t be reasonable to write: lv = new Vehicle();

If this was allowed, we could send any LandVehicle message to the Vehicle, even though some of the LandVehicle messages won’t be understood by Vehicle.

4.3.2 Polymorphic Messages Any message, in a pure object-oriented language, can have more than one method associated with it. This happens either because the methods appear independently on more than one

83

84

Chapter 4

class or because a method is redeﬁned by subclasses. Redeﬁned methods normally have similar algorithms, but methods deﬁned independently usually have completely different algorithms. For example, consider the four classes in Figure 4.5, where any possible relationships between humans and birds have been ignored. All of these classes have a message ﬂyTo, which means ‘Use ﬂight to get to the given location’: the method is abstract on Bird and concrete on the other three classes. Since ﬂyTo has three implementations, it is a polymorphic message.

Bird flyTo(:Location) Human flyTo(:Location)

Finch

Eagle

flyTo(:Location)

flyTo(:Location)

Figure 4.5: Polymorphic animal messages

Although ﬂyTo on Human can be compared to ﬂyTo on Bird, the meanings are really quite different: for birds, ﬂying involves ﬂapping and gliding; for humans, it involves getting on a plane. Polymorphic messages are linked to inheritance, in the same way as polymorphic variables. For example, in the following code fragment, we would expect the run-time system to execute the ﬂyTo method on Finch, not the one on Bird (which is abstract anyway, so there would be nothing to execute): Bird b = new Finch(); b.flyTo(someLocation);

Here we see the full power of abstract methods: we’ve been able to state that ‘All birds have a ﬂyTo message’, even though we can’t deﬁne one at the level of Bird – the run-time system has the job of picking the correct implementation. If we added the following code fragment to the one above, we would be telling an Eagle to ﬂy to someLocation, which presumably involves more of the gliding and less of the ﬂapping. b = new Eagle(); b.flyTo(someLocation);

Dynamic Binding

A polymorphic variable is just a place holder – the business end is the object it happens to be pointing to. So, in the following code fragment, we have one Cat, but two references to it (each reference is a separate name for the cat): Cat tiddles, tom; tiddles = new Cat("Hfrrr"); tom = tiddles;

4.4 DYNAMIC BINDING Dynamic binding means attaching a message to a method at run time. This is the way that object-oriented languages cope with polymorphic variables and redeﬁned methods. Take a look at Figure 4.6. Here, we have two abstract classes – Shape and Quadrilateral – and one concrete class – Square. All three classes have a getPerimeter message: on Shape, it is abstract; on Quadrilateral, it sums the length of all four sides; on Square, it multiplies the length of a single side by four. The fact that aSquare is an instance of class Square has been shown as a UML dependency, a dashed, open-ended arrow.

Shape

fCo

rne

rs(

)

getPerimeter():int getNumberOfCorners():int

mb

er

getPerimeter():int getNumberOfCorners():int

()

Nu

im

et

Quadrilateral

get

tP er

erO

ge

aSquare

dependency

Square getPerimeter():int

Figure 4.6: Dynamic binding and redeﬁnition

85

86

Chapter 4

In the following code fragment, how does the run-time system know which of the three methods to execute when the getPerimeter message is sent? Shape sh = new Square(); int i = sh.getPerimeter();

Conceptually, what happens is this: the object that receives the message (a Square in this case), knows its own class, so it goes there to look for the method; since Square deﬁnes a getPerimeter method, it is executed. Figure 4.6 also shows a getNumberOfCorners message, which is abstract on Shape and concrete on Quadrilateral. In the following code fragment, the Square will again look to its class for a getNumberOfCorners method. int j = sh.getNumberOfCorners();

This time, there is no matching method, so the search continues with the Quadrilateral superclass. Quadrilateral does have a matching method, so that’s the one that is executed. Although this dynamic binding algorithm looks a little slow, it is only a conceptual algorithm and can be implemented much more efﬁciently, especially with the help of a static type system. With static typing, we can be sure that dynamic binding will ﬁnd a concrete method somewhere in the superclass chain (because the type of every variable has to be declared and we can’t create abstract objects). With a dynamic type system, we might encounter an abstract method or fall off the top of the hierarchy altogether – this would result in a run-time error message. For multiple inheritance, dynamic binding is more complicated, but it is still feasible and is not appreciably slower than with single inheritance. Figure 4.7 shows our Shape hierarchy with an additional class Triangle. Consider the following code fragment: Shape sh; Triangle tr = new Triangle(); Square sq = new Square();

Which of the following assignments would be correct? (Remember to think ‘Is it a kind of?’) 1. sh = tr; 2. sh = sq; 3. sq = tr; 4. tr = sq; 5. tr = sh; 6. sq = sh;

ACTIVITY 2

Polymorphism Guideline

Shape getPerimeter():int

Triangle getPerimeter():int getHeight():int

Quadrilateral getPerimeter():int

Square getPerimeter():int

Figure 4.7: Polymorphism and type systems

ACTIVITY 3

If we attach sh to the square with: sh = sq;

What do you think the effect of the following message sends would be (assuming a static type system)? 1. sh.getPerimeter(); 2. sq.getPerimeter(); 3. tr.getPerimeter(); 4. tr.getHeight(); 5. sq.getHeight(); 6. sh.getHeight();

4.5 POLYMORPHISM GUIDELINE Before we leave the subject of polymorphism, consider the following style guideline: ‘Always program using as high a level of abstraction as you can’. That is, always declare the type of

87

88

Chapter 4

your ﬁelds, local variables and method parameters to be the highest class possible in the inheritance hierarchy, then let polymorphism do the rest. The reason for this guideline is that the higher the abstraction level you use, the more reusable your code will be. For example, any code written in terms of Shape will work for any kind of Shape – Square, Quadrilateral, Triangle, and any other subclasses that we might add later. But code written for Square will only work for Square objects, code written for Triangle will only work for Triangle objects, and so on. Designers of class hierarchies have a complementary responsibility here: they should make sure, wherever possible, that the speciﬁcs of each class are hidden behind general messages that apply to all the related classes. For example, drawing a square on screen is signiﬁcantly different to drawing a triangle. Rather than exposing the difference to the client programmer, we should design a message, or a group of messages, that can be placed on Shape. For example, we might provide a draw message on Shape, so that the client programmer can tell a Shape to draw itself without knowing anything about how the drawing is done. Alternatively, we could add a message called getLines to the Shape class, so that a client programmer could retrieve the lines that make up the shape’s perimeter and draw them on screen, without caring about how they were generated.

4.6 TYPE CASTING In a statically typed language, when we pass a value from one context to another we need to be sure that the new context is compatible with the old one (for example, we don’t want to pass a Banana to a context that expects an Employee). Even when we’re sure that the new context is compatible with the old one, we may still need to convert the value to a different type (for example, passing an Employee object from an Employee context to a Person context). There are three situations in which a value changes context: • Expression evaluation: In the expression 2 + 2, the compiler is being asked to add two integers together to produce another integer. No problem, because the new context and the old context are both integer. On the other hand, if we code the expression 3.75 + 2, the compiler begins with a real number and an integer, but it must produce a real number. In this case the integer is compatible with the new context – because we can add any two types of number together to produce another type of number – so the expression is valid. However, computers represent integers and real numbers in different ways so, in order to perform the arithmetic above, the compiler must ﬁrst convert the integer into a real number (because integers can be represented as real numbers, but not vice versa).

Explicit Casts

Some languages allow objects to take part in expressions, as in the Java expression ‘‘The These expressions must obey similar rules to the primitive examples above. In this case, the compiler will translate the Date into a String by sending it the toString message. • Assignment: An assignment that doesn’t require a conversion is Person p = new Person(), because the expression context and the variable context are both Person. However, the assignment Person p = new Employee() requires a conversion from Employee to Person. Person pointers are not represented differently from Employee pointers but the compiler must be sure that an Employee can be used as a Person. If Person is a direct or indirect superclass of Employee, the conversion will be safe, because the client programmer can only use elements of the Person class. An example of a primitive assignment that requires a conversion is ﬂoat f = 2, in which the integer 2 must be converted to a ﬂoating point number. • Parameter passing: This is the same as assignment. The value of the actual parameter is assigned to the declared parameter, for example, aUniversity.enrollPerson(new Employee()). A primitive example is aLiquid.setBoilingPoint(100), where the parameter is declared to be ﬂoat. date is ” + aDate.

Converting a value from one type to another is called casting (because we cast the value in a new light). The examples above are all implicit casts because the client programmer doesn’t have to do anything special: the compiler can see that the conversions are valid and just does them. In general, an implicit cast is possible if the new context is wider than the old context. In terms of primitive values, wider means ‘the new context can accommodate all possible values of the old context’; for object values, wider means ‘the new context is a direct or indirect superclass of the old context’. Dynamically-typed languages have a much easier time passing values between contexts. For example, when executing the Smalltalk method addEmployee with a Banana as a parameter, no compatibility testing or casting needs to be done: as long as addEmployee only uses elements that are common to Employee and Banana, the method will run without error. (We would still have a problem later if we attempted to transfer money to the Banana’s bank account, but that would manifest itself as a run-time error.)

4.7 EXPLICIT CASTS Implicit casts enrich statically-typed programming languages by allowing us to combine different types of values in expressions and by allowing us to assign different types of values to variables. But we can go one step further, by allowing the programmer to use an explicit cast, to move from one context to a compatible, but narrower, context.

89

90

Chapter 4

It turns out that the explicit casting of a value can cause problems, but the explicit casting of a pointer is not nearly so bad. This is because, when we cast a value, we’re modifying the way the value is stored – storing the value using a narrower representation may lose information. In contrast, if we cast a pointer, we’re modifying the way we access the value, but the value itself remains unchanged. In a pure object-oriented language, primitives are always accessed as values and objects are always accessed via pointers (efﬁciency versus polymorphism). In contrast, in a language like C++, primitives and objects can be accessed either as values or via pointers: the programmer has the choice. The upshot of this is that C++ programmers have to be much more careful. For the rest of this discussion, let’s assume that we’re using a pure object-oriented language. Let’s also assume that explicit casts are used only when necessary: turning an implicit cast into an explicit one is allowed, but it has no effect. Java allows a programmer to write int i = (int) 3.75 (int) explicitly

casts the real value 3.75 as an integer. The programmer, aware that an implicit cast from ﬂoat to int is impossible, is telling the compiler to force a square peg into a round hole. In real life, forcing a square peg into a round hole tends to shear off the peg’s corners. In programming, it’s no different: the fractional part of the real number is sheared off. The explicit cast above only works at all because int and ﬂoat are compatible: boolean b = (boolean) 3.75 would never make sense. In general, if there exists an implicit cast in one direction, we can force an explicit cast in the other direction. How about explicit casting for object pointers? Could the following example ever work? Employee e = (Employee) new Person();

For pure object-oriented languages, the answer is no. Remember that casting a pointer doesn’t change the representation of the value, it just changes the way we look at it. So, our example is asking the compiler to let us treat a Person as an Employee. If the compiler allowed this, we might try to use an Employee-only element which wasn’t there (getPayrollNumber, for example). How about the following pair of statements, could these ever work? Person p = new Employee(); Employee e = (Employee) p;

In this case, the object we’re trying to get e to point to really is an Employee, so the explicit cast ought to be safe – we can’t abuse an Employee via an Employee pointer. But the compiler is stuck. In order to work out that those statements are valid, the compiler would have to analyze both statements. As already noted, compilers don’t do this kind of analysis because it’s difﬁcult and often impossible.

Genericity with Templates

However, occasionally this is a useful thing for the programmer to be able to do. Therefore, the compiler enlists the help of the run-time system in order to allow the explicit cast, but only because Employee is a subclass of Person. Subsequently, the run-time system needs to check that the object pointed to by p is an instance of Employee at the time of the assignment (or an instance of a direct or indirect subclass): if not, a run-time error will occur. Because the explicit cast of an object pointer involves the run-time system, it is sometimes referred to as a dynamic cast. Some people also use the terms upcast and downcast to refer to implicit and explicit casting of object pointers, respectively, because an implicit cast moves the type up the hierarchy and an explicit casts moves the type down. Downcasts aren’t needed very often. You should only use them when retrieving an object using a generic message – a message whose return type must be compatible with all possible objects. For example, the Stack class we saw earlier has a generic message for popping an object: it always returns a pointer of type Object, regardless of what’s actually on the top. (It also has a generic message for pushing an object: the parameter’s type is Object, which might require an upcast.) Generic messages allow the stack to work for all kinds of object – cars, strings or whatever. But they wouldn’t allow a client programmer to do the following: aStack.push(new Plate("Domestic blue")); Plate aPlate = aStack.pop();

The ﬁrst line works, because it’s an implicit upcast, but the second line doesn’t. However, with the help of a downcast, the client programmer can do this: aStack.push(new Plate("Domestic blue")); Plate aPlate = (Plate) aStack.pop();

4.8 GENERICITY WITH TEMPLATES Some languages, such as Java, hardly need downcasts at all, because they have a facility called genericity, which is better than downcasts. A generic class uses one or more class parameters to refer to the types of objects that it expects to deal with. In Java, a generic class is called a template. For example, the Java programmer can deﬁne a Stack class like this: public class Stack<X> { private List<X> list; public Stack<X>() { list = new LinkedList<X>(); } public void push(X anX) {

91

92

Chapter 4

list.addElement(anX); } public X peek() { return list.lastElement(); } public boolean isEmpty() { return list.numberOfElements() == 0; } public Object pop() { X anX = list.lastElement(); list.removeLastElement(); return anX; } }

Despite the new syntax, you should be able to pick out the important parts of this class. The most important aspect is the class referred to throughout as X: this is a place-holder for the actual kind of Stack created by the client programmer. Notice also that the type of the ﬁeld is more general than the implementation class (this is just good style for object-oriented code in general). Now, the client programmer can manipulate a stack using the following lines of code: Stack s = new Stack(); s.push(new Plate("Hospital white")); Plate p = s.pop();

Because the programmer has declared that the Stack contains Plate objects, there’s no need for an explicit cast before s.pop(). A programmer who wants a quick-trick-brick stack can write the following: Stack s = new Stack();

The Stack class is a relatively simple generic type, because none of its methods needs to use the elements of the objects it contains. If we want to make assumptions about the class of the stacked objects, we need something more. For example, we might decide that Stack objects should only contain food – this would allow the Stack methods to treat the pushed objects as Food, counting calories, for example. If we want to make assumptions about the class parameters, we have to use constrained genericity: adding constraints to the class parameters. Here’s part of a modiﬁed Stack class written in Java, one that only works for different kinds of Food: public class Stack<X extends Food> { ... private int caloriesConsumed; ...

Review Questions

public X pop() { X anX = list.lastElement(); list.removeLastElement(); // Now add anX’s calories to the total, // only possible because X is always a Food caloriesConsumed = caloriesConsumed + anX.getCalories(); return anX; } ... }

4.9 SUMMARY In this chapter, we looked at: • Type systems, which stop us misusing values by forcing us to declare how we intend to use a value. A static type system detects abuses at compile time while a dynamic type system waits until run time. • Polymorphism, which enables a variable to hold different types of value and a message to be associated with more than one method. The speciﬁc type of value or method applicable in any case is determined at run time. • Casting between object types: with implicit casting, the compiler can automatically convert between types of variable; with explicit casting, the programmer must specify that an object is to be considered as a different type. • Templates in Java, in which a generic class uses parameters to refer to the types of objects that it expects to deal with.

FURTHER READING For more on Java templates, a relatively recent addition to the language, go to the Java web site at java.sun.com.

REVIEW QUESTIONS 1. With reference to Figure 4.8, which methods correspond to the following message sends (in the order given)? Choose only one option. tr.height(); sh.perimeter(); sq.height(); sq.perimeter(); sh.height(); tr.perimeter();

93

94

Chapter 4

(a) (b) (c) (d)

3, 1, none (error), 4, none (error), 5 3, 5, none (error), 4, 3, 5 3, 1, none (error), 4, 3, 5 3, 5, none (error), 4, none (error), 5

Shape

2: perimeter()

Rectangle

1: perimeter()

Triangle 3: height()

4: perimeter()

5: perimeter()

Square

Shape sh; Triangle tr = new Triangle(); Square sq = new Square(); sh = tr; Figure 4.8: For use with Review Question 1

2. With reference to Figure 4.8, which of the following message sends would be allowed by a compiler? Choose all options that apply. (a) (b) (c) (d) (e) (f)

sh.perimeter(); tr.perimeter(); sh.height(); sq.height(); sq.perimeter(); tr.height();

3. With reference to Figure 4.8, which of the following assignments would be allowed by a compiler? Choose all options that apply. (a) (b) (c) (d)

sq = sh; sh = tr; tr = sq; tr = sh;

Answers to Activity 3

(e) sh = sq; (f) sq = tr; 4. What can an object-oriented type system be used for? Choose only one option. (a) (b) (c) (d) (e) (f) (g)

Improving runtime performance. Preventing misuse of a class. Avoiding spelling mistakes. Making sure that all messages invoke a concrete method. Documentation. All of the above. None of the above.

5. What does the term ‘polymorphism’ refer to? Choose all options that apply. (a) The ability of a variable to point at different classes of object at different times. (b) The fact that a message with the same signature can invoke different methods at different times. (c) All object-oriented programming languages are different. (d) All object-oriented methodologies use a different notation. 6. What is a generic class? Choose only one option. (a) A class with no declared copyright. (b) A class that represents all objects. (c) A class that has other classes as parameters.

ANSWERS TO ACTIVITY 2 1. This is ﬁne: a Shape variable must be able to point at a Triangle, because a Triangle is a kind of Shape. 2. This is ﬁne: a Square is a kind of Shape. 3. This is an attempt to point at a Triangle with a Square variable, so that must be wrong, because a Triangle is not a kind of Square. 4. Similarly, a Square is not a kind of Triangle. 5. A Shape is not a kind of Triangle. 6. A Shape is not a kind of Square.

ANSWERS TO ACTIVITY 3 1. The getPerimeter message is introduced on Shape, so that is correct; since sh is pointing at a Square, it’s the getPerimeter method on Square that is invoked.

95

96

Chapter 4

2. 3. 4. 5.

The getPerimeter method on Square is invoked. The getPerimeter method on Triangle is invoked. The compiler is happy and the getHeight method on Triangle is executed. This is an error, because there is no getHeight message on Square, nor on its superclasses – the compiler will reject this one. 6. We’re trying to send getHeight to a Shape variable. If sh points at a Triangle, the message would not be a problem; but sh could point at a Square, in which case the message would be invalid. Compilers for static type systems tend to err on the side of caution: as the message might arrive at an object that doesn’t know how to deal with it, the compiler would reject the statement.

ANSWERS TO REVIEW QUESTIONS 1. In Figure 4.8, the methods correspond to the message sends d. 3, 5, none (error), 4, none (error), 5 2. In Figure 4.8, the following message sends would be allowed by a compiler: (a) (b) (e) (f)

sh.perimeter(); tr.perimeter(); sq.perimeter(); tr.height();

3. With reference to Figure 4.8, the following assignments would be allowed by a compiler: (b) sh = tr; (e) sh = sq; 4. An object-oriented type system can be used for f. All of the above: (a) (b) (c) (d) (e)

Improving runtime performance. Preventing misuse of a class. Avoiding spelling mistakes. Making sure that all messages invoke a concrete method. Documentation.

5. The term ‘polymorphism’ refers to a. The ability of a variable to point at different classes of object at different times and b. The fact that a message with the same signature can invoke different methods at different times. 6. A generic class is c. A class that has other classes as parameters.

5

Software Development Methodologies

I

n this chapter, we’re going to take a look at the software development process itself: the steps we go through to write good software and the things that we produce along the way. Collectively, the process, its steps and its products are referred to as a methodology.

Learning Objectives Understand the classical phases of software production.

•

Compare static (structural) modeling with dynamic (time-based) modeling.

Understand the notation of Uniﬁed Modeling Language (UML).

Chapter Outline 5.1 Introduction 5.2 Classical Phases in Software Production 5.3 Software Engineering and the Waterfall Methodology 5.4 New Methodologies 5.5 Object-Oriented Methodologies 5.6 Ripple Overview 5.7 Summary Further Reading Review Questions Answers to Review Questions

5

100

Chapter 5

5.1 INTRODUCTION All software, especially large pieces of software produced by many people, should be produced using some kind of methodology. Even small pieces of software developed by one person can be improved by keeping a methodology in mind. A methodology is a systematic way of doing things. It is a repeatable process that we can follow from the earliest stages of software development (the germ of an idea or a new business opportunity) through to the maintenance of an installed system. As well as the process, a methodology should specify what we’re expected to produce as we follow the process (and what form the products should take). A methodology will also include advice or techniques for resource management, planning, scheduling and other management tasks. Good, widely available methodologies are essential for a mature software industry – the alternatives are highly unsatisfactory. The worst alternative is downright chaos, where members of the development team run around in a panic, wondering how on earth they’re going to come up with the latest system that they’ve been told to implement. Only slightly better is the situation where an ad hoc development process is designed by amateur methodologists within one organization – such in-house efforts need to be learnt by every newcomer to the organization and they’re useless to anyone moving on. Although most methodologies are designed to cope with teams of developers producing large amounts of software, understanding the basics of a good methodology is essential for those at the other end of the scale too (lone developers working on small problems) and at all points in between. This is because: • A methodology can help to impose discipline on the coding effort. • Going through even the basic steps of a methodology increases our understanding of the problem, improving the quality of our solution. • Writing lines of code is only one of the many activities in software development: performing some of the other activities helps us to spot conceptual and practical mistakes before we commit them to source code. • At every stage, a methodology speciﬁes what we should do next, so we’re not left scratching our heads, thinking ‘Okay, what now?’ • A methodology helps us to produce code that is more extensible (easier to change), more reusable (applicable to other problems) and easier to debug (because it has more documentation). Large development projects also beneﬁt from: • Documentation: All methodologies promote thorough documentation of every stage of the development effort, so that the ﬁnished system is not an impenetrable monolith.

Introduction

• Reduced latency: Since the workﬂows, activities, roles and inter-dependencies are better understood, there is less opportunity for human (and other) resources to lie idle for want of something to do. • Improved chances of delivery on time and within budget. • Better communication between users, sales people, managers and developers: A good methodology is based on logic and common sense, so it will be easy for all participants to grasp the basics; thus, we have a more orderly development, with less scope for misunderstanding and wasted effort. • Repeatability: Since we have well-deﬁned activities, similar projects should be delivered to similar time-scales and with similar costs. If we produce similar systems over and over again for different customers (e-commerce shop fronts, for example) we can streamline the methodology in order to concentrate solely on the unique aspects of the latest development; eventually we might automate parts of the development and even sell the automations to third parties (think of a ‘shop front in a box’ product). • More accurate costing: When asked ‘How much will it cost?’, there will be less temptation to reply ‘How much have you got?’ A good methodology will address at least the following issues: • Planning: Deciding what needs to be done. • Scheduling: Mapping out when things will be done. • Resourcing: Estimating and acquiring the human, software, hardware and other resources that are needed. • Workﬂows: The subprocesses within the wider development effort (for example, designing the system architecture, modeling the problem domain and planning the development effort). • Activities: Individual tasks within a workﬂow, such as testing a component, drawing a class diagram or detailing a use case, too small or indeﬁnable to be a workﬂow in their own right. • Roles: The parts played by personnel within the methodology (developer, tester or sales person). • Artifacts: The products of the development effort: pieces of software, design documents, training plans and manuals. • Education: Deciding how to train personnel, if necessary, to fulﬁll their required roles; deciding how end users (staff, customers, sales people) will learn how to use the new system. For the purposes of this book, we won’t be looking at the details of an industrial methodology – that would require a book in its own right. Instead, we’ll use a specialpurpose methodology, called Ripple, which is derived from but is rather simpler than

101

102

Chapter 5

the Rational Uniﬁed Process. Before we examine Ripple, we need to have an idea of the processes, activities and artifacts of software development in general.

5.2 CLASSICAL PHASES IN SOFTWARE PRODUCTION So, what does software development involve? There are a number of phases common to every development, regardless of methodology, starting with requirements capture and ending with maintenance. With the traditional approach, you’re expected to move forward gracefully from one phase to the other. With the modern approach, on the other hand, you’re allowed to perform each phase more than once and in any order. The list below describes the common phases in software development – you may have seen different names for some of these, but the essentials remain the same. At this stage, we’re interested in the intent of the phases rather than details of how you might actually go about performing them. Be warned, though, that some methodologists combine requirements and analysis, while others combine analysis and design.

5.2.1 Requirements Requirements capture is about discovering what we’re going to achieve with our new piece of software and has two aspects. Business modeling involves understanding the context in which our software will operate – if we don’t understand the context, we have little chance of producing something to enhance that context. The sort of question we ask during the business modeling phase is ‘How does a customer purchase a television from this shop?’ System requirements modeling (or functional speciﬁcation) means deciding what capabilities the new software will have and writing down those capabilities. We need to be clear about what our software will do and what it won’t do, so that the development doesn’t veer off into irrelevant areas and we know both when we’ve ﬁnished and whether we’ve been successful. The sort of question we ask during the system requirements modeling phase is ‘How do we update the inventory system when a television has been purchased?’

5.2.2 Analysis Analysis means understanding what we’re dealing with. Before we can design a solution, we need to be clear about the relevant entities, their properties and their inter-relationships. We also need to be able to verify our understanding. This can involve customers and end users, since they’re likely to be subject-matter experts. The sorts of question we ask during the

Classical Phases in Software Production

analysis phase are ‘What products do we sell in this shop? Where do they come from? How much do they cost?’

5.2.3 Design In the design phase, we work out how to solve the problem. In other words, we make decisions, based on experience, estimation and intuition, about what software we will write and how we will deploy it. System design breaks the system down into logical subsystems (processes) and physical subsystems (computers and networks), decides how machines will communicate, chooses the right technologies for the job, and so on. The sort of decision we make during the system design phase is ‘We’re going to use an intranet and the Java Messaging Service for communicating sales results to head ofﬁce.’ In subsystem design we decide how to cut each logical subsystem into effective, efﬁcient and feasible code. The sort of decision we make during the subsystem design phase is ‘Line items in an inventory are implemented as a hash table, keyed by part number.’

5.2.4 Speciﬁcation Speciﬁcation is an often-ignored, or at least often-neglected, phase. The term speciﬁcation is used in different ways by different developers. For example, the output of the requirements phase is a speciﬁcation of what the system must be able to do; the output of analysis is a speciﬁcation of what we’re dealing with; and so on. In this book, the term is used to mean ‘describing the expected behavior of our programming components’. (Since the speciﬁcation techniques described are performed on classes of objects, some of the confusion can be avoided by using the term class speciﬁcation.) A class speciﬁcation is a clear, unambiguous description of the way the components of our software should be used and how they will behave if used properly. The sort of statement we make during the speciﬁcation phase is ‘If the shop assistant object is logged on, it can ask the store object for today’s special offers; in return, it receives a list of products, sorted in alphabetical order’. This book gives special attention to speciﬁcation, because of the crucial underlying principle of Design by Contract. The idea behind a contract is that whenever one piece of software calls upon the services of another, both the caller and the called have obligations to fulﬁll. Bearing software contracts in mind is useful at all stages of development. Speciﬁcation can be used in the following ways: • As a basis for designing test software to exercise the system. • To demonstrate that our software is correct (this is desirable for life-critical applications). • To document our software components to the extent that they could be implemented by third parties. • To describe how our code can be reused safely by other applications.

103

104

Chapter 5

5.2.5 Implementation This is where we do the donkey work, writing pieces of code that work together to form subsystems, which in turn collaborate to form the whole system. The sort of task we carry out during the implementation phase is ‘Write the method bodies for the Inventory class, in such a way that they conform to their speciﬁcation’. Although we would expect most of the difﬁcult coding decisions to have been made before we reach this phase (during design), there is still plenty of scope for creativity: although the public interfaces of our software components will have been well designed, speciﬁed and documented, programmers have free rein to decide on the inner workings. As long as the end result is effective and correct, everyone will be happy.

5.2.6 Testing When our software is complete, it must be tested against the system requirements to see if it ﬁts the original goals. The sort of question we ask during the testing phase is ‘Can a shop assistant use the till interface to sell a toaster, decreasing the product’s inventory as a side-effect?’ As well as this kind of conformance testing, it’s a good idea to see if our software can be broken via its external interfaces – this helps to protect us against accidental or malicious abuse of the system when it’s been deployed. It is a good idea for programmers to perform small tests as they go along, to improve the quality of the code that they deliver. Generally speaking, however, major tests should not be designed, implemented or carried out by the developers who wrote the software. To understand why, consider buying a new house and spending vast amounts of time and money refurbishing it from top to bottom. It’s unlikely that you would want to whack the structures and ﬁxtures with a sledgehammer to see if they’re durable, ask passing strangers whether they think that you have good taste or pretend to be a burglar to see if you can break in. These are exactly the kinds of things that we need to be doing during software testing. (It helps if members of the test team have a cruel streak.)

5.2.7 Deployment In the deployment phase, we’re concerned with getting the hardware and software to the end users, along with manuals and training materials. This may be a complex process, involving a gradual, planned transition from the old way of working to the new. The sort of task we carry out during the deployment phase is ‘Run the program setup.exe on each server machine and follow the instructions that appear’.

5.2.8 Maintenance When our system is deployed, it has only just been born. A long life stretches before it, during which it has to stand up to everyday use – this is where the real testing happens.

Classical Phases in Software Production

The sort of problem we discover during the maintenance phase is ‘When the log-on window opens, it still contains the last password entered.’ As software developers, we’re normally interested in maintenance because of the faults (bugs) that are found in our software. We must ﬁnd the faults and remove them as quickly as possible, rolling out ﬁxed versions of the software to keep the end users happy. As well as faults, our users may discover deﬁciencies (things that the system should do but doesn’t) and extra requirements (things that would improve the system). From the business point of view, we would hope to ﬁx and improve our software over time to maintain competitive advantage.

5.2.9 Key Questions These key questions will help you to remember the purpose of each of the software development phases: • Requirements phase: ‘What is our context?’ ‘What are we trying to achieve?’ • Analysis phase: ‘What entities are we dealing with?’ ‘How can we be sure we have the right ones?’ • System design phase: ‘How are we going to solve the problem?’ ‘What hardware and software will we need in the ﬁnished system?’ • Subsystem design phase: ‘How are we going to implement the solution?’ ‘What will the source code and supporting ﬁles look like?’ • Speciﬁcation phase: ‘What rules govern the interfaces between the system components?’ ‘Can we remove ambiguity and ensure correctness?’ • Implementation phase: ‘How can we code the components to meet the speciﬁcation?’ ‘How do we write stylish code?’ • Testing phase: ‘Does the ﬁnished system satisfy the requirements?’ ‘Can we break the system?’ • Deployment phase: ‘What do the system administrators have to do?’ ‘How can we educate the end users?’

105

106

Chapter 5

• Maintenance phase: ‘Can we ﬁnd and ﬁx the faults?’ ‘Can we improve the system?’

5.3 SOFTWARE ENGINEERING AND THE WATERFALL METHODOLOGY During the 1970s, software gurus gave much thought to the problem of how best to write software. How could they replace ad hoc, proprietary mechanisms with scalable, portable methodologies? What the gurus came up with was software engineering. The idea was that software production could be like building a real-world structure, such as a road bridge. With the help of physics, engineering is systematic: if we follow the rules, we will deliver a working product, complete with safety margins to protect against abnormal conditions. There are some obvious drawbacks to this analogy. Most programming is done with imperative programming languages that require the programmer to tell the computer exactly what to do – statement by statement, branch by branch, function by function. This is analogous to the bridge engineer having to tell a piece of steel exactly how to behave, rather than relying on the laws of physics. And that’s not all: the programmer also has to assemble the data piece by piece. Can you imagine a bridge builder having to assemble the bridge’s road surface, stone by stone, and then telling each stone exactly which other stones it was touching? As well as requiring the programmer to be precise about behavior and structure, imperative languages don’t cope well with imprecise data. Generally speaking, if a piece of data in the system is slightly inaccurate, the system might not behave as expected. In contrast, a bridge builder can happily connect girders using rivets that are all slightly different in size: the engineer can count on tolerances and margins of error, but the programmer can’t. These drawbacks did not prevent the growth in popularity of software engineering and methodologies grew up around the assumption that software production could be systematic and predictable. This led to the so-called waterfall methodology (see Figure 5.1). Development ﬂows smoothly over the classical phases (requirements, analysis, system design, etc.), with each phase being completed satisfactorily before the next phase is attempted. It’s easy to plan (because the plan is similar every time) and it’s easy to schedule (use the complexity of the problem and the number of developers to work out how long the development will take, then divide the result by the total number of phases...). The waterfall methodology allows us to have developers with different kinds of expertise at each stage (hence the classical roles of business analyst, systems analyst, designer, programmer, tester and system administrator). Each team of specialists slogs away during their own phase, until

Software Engineering and the Waterfall Methodology

Requirements Analysis Design Specification Implementation Testing Deployment Maintenance

Figure 5.1: Waterfall development

they’re sure that they have solved their part of the problem; then they document their work, using their own jargon and notation, and pass the baton to the next team of specialists. The waterfall methodology is a nice idea, but unrealistic. Even if we’re clever enough to work out how long the development might take, before we’ve looked in detail at the problem, we can’t tell what difﬁculties we will encounter along the way (bad design decisions, pernicious faults, inadequate technology, or earthquakes). So, any individual phase may take longer than expected. Also, work tends to expand to ﬁll the time available, so the phases that precede a problem are likely to have used up all the time available to them. The net effect is that the whole project is delivered late. In practice, this is exactly what happens in the majority of cases. The waterfall methodology may fail for other reasons too, such as analysis paralysis – the analysts are reluctant to sign off their documentation because they can’t decide whether they’ve understood and documented the system entities well enough to allow the designers to do their job. To be fair, this kind of problem is not restricted to analysts: the designers might be worried that their design is inadequate; the speciﬁers might worry that their speciﬁcation is too ambiguous to be coded; and so on, leading to even more delays. In practice, it’s not possible to complete each phase perfectly. Throughout the development, members of the team will discover problems with the work that’s gone before. Whenever this happens, we have a tricky choice: we can return to the earlier documentation and ﬁx it, but that would mean climbing back up the waterfall (which we’re not supposed to do); or we could make a note of the problems, with the intention of repairing the documentation at the end of the project (which rarely happens, so the ﬁnal documentation doesn’t match the ﬁnal system). And what about the end users? Are they going to get what they want or need? The potential users of the system will be colleagues or third parties, probably paying us real

107

108

Chapter 5

money. These customers will presumably have been consulted during the requirements phase: we will have asked them how they currently work; we will have brainstormed the kind of system functions that could be delivered; we will have reached an understanding about what is going to be delivered and what isn’t. But what if we ﬁnd that we can’t deliver some of the functions, because they’re too difﬁcult or we run out of time? The customers won’t ﬁnd out until testing or maybe even deployment – far too late to change anything. What if the project takes two years to complete? End users’ requirements will normally change signiﬁcantly in two years. Are we going to deliver a system that’s no longer relevant? What we need is a way of involving users throughout the development, so that we don’t deliver something that’s a nasty surprise. We also need to reduce the time between the promised functionality and the point where we can demonstrate that functionality. But the waterfall model is too rigid to allow us to incorporate user feedback and take corrective action. And these aren’t the only problems. For example, the waterfall methodology will always suffer from being focussed on solving one particular problem – this makes it more difﬁcult to produce reusable code. Eating an elephant is a daunting task (see Figure 5.2); if you were told to eat one, would you prefer to open wide and swallow the beast whole or to eat a manageable portion, have a rest, eat another portion, have a rest, and so on, until the whole elephant had been consumed? Waterfall development uses the ﬁrst approach and attempts to produce a complete solution in one go. As we’ll see, it’s possible to use the second approach, delivering the solution piece by piece. Given the large number of problems with the waterfall methodology, some of which should have been obvious at the outset, you may ﬁnd it hard to believe that any software development was ever done this way. Well, it was and in some places it still is. Some corporations have been quite happy to base large developments on the waterfall methodology, integrate it into their software departments as best practice and base entire career ladders on it (start as programmer, get promoted to designer, then systems analyst, and so on). However, most object-oriented enthusiasts and, shall we say, enlightened corporations, prefer something rather more ﬂexible. Having debunked the waterfall methodology, it’s only fair to point out that it is still useful in the following cases: • When repeating a particular kind of development with only minor differences (for example, an e-commerce shop front for a particular company may only differ from the previous one in terms of product descriptions, prices, company name and logo). • As a framework for learning the different techniques used in software development: although the waterfall methodology is too simplistic for real-world development, it does contain the classical phases in a logical order, so it’s good for learning. • As a single pass around a spiral methodology.

Software Engineering and the Waterfall Methodology

Figure 5.2: Eating an elephant

• As a framework to support an iterative methodology. • For the quick development of small projects with small numbers of developers, as in prototyping, production prototyping, proof-of-concept or Rapid Application Development (RAD). Four new styles of programming have been developed as a result of the simplicity and power of objects, the reuse of code in new applications, and the advent of application builders: • A software prototype, just like an engineering prototype, is something that we build in order to try out some of the functions of the ﬁnished product. A prototype doesn’t need to be elegant, or industrial strength, because it’s just an experiment. We should set aside a prototype once it has served its purpose and start afresh. • A production prototype is similar to a prototype, except that we retain some or all of the code through to completion of the project. • A proof-of-concept is a project or a piece of software designed to demonstrate the feasibility of some technology, or group of technologies. For example, we might need to convince a

109

110

Chapter 5

customer that we’re qualiﬁed to take on a particular project or we might need to convince management to adopt a new approach to software production. • The phrase Rapid Application Development (RAD) was coined by enthusiasts of object orientation to mean building a system more quickly than with traditional techniques. As object-oriented systems became practical in the 1980s, object enthusiasts were able to impress traditional developers (and managers) with how quickly they could assemble small systems. An application builder is a tool that allows the programmer to assemble software in the same way that a computer manufacturer assembles hardware. As time has passed, application builders have employed larger and larger components so that systems can be constructed more quickly than before. In most cases, objects destined for an application builder need to have a particular kind of interface in order to work properly. This means that the programmer needs to follow predeﬁned style guidelines rather than designing their objects any way they see ﬁt. For example, Java application builders expect to work with objects that follow the JavaBeans guidelines [Campione et al. 98].

5.4 NEW METHODOLOGIES Before we can replace the waterfall methodology with something better, we need to accept that it is impossible to develop a piece of software in one pass. However hard we try, the ﬁrst time around our software will be incomplete or imperfect or both. Therefore, we need to perform the classical phases of software development several times, adding to and perfecting the system as we go.

5.4.1 Spiral Methodology One way to look at this is as a spiral methodology (see Figure 5.3). We start, as ever, with requirements capture, which may be relatively complete or rather vague at this stage; next, we perform some exploratory analysis to increase our understanding of what it is that we’re dealing with; then we sketch out a system design that we feel will ﬁt the requirements and design part of the system; then, despite the fact that all the preceding phases are incomplete, we write some code. Once we’ve ﬁnished our initial coding effort we can try out what we have so far, perhaps by running some informal tests or by showing what we have to our sponsors (end users, managers and customers paying for the system). By the time we’ve been through the cycle once, we’ve increased our understanding of the problem domain and our understanding of the proposed solution. We’ve also involved our sponsors, so that they can correct any misunderstanding of the business or the functionality that they expect to see in the ﬁnal system. Armed with our new body of knowledge, we

New Methodologies

Specification

Design

Analysis

Implementation

Testing

Deployment

Maintenance

Requirements

Figure 5.3: Spiral development methodology

can go around the spiral again: now we ﬂesh out the requirements; we’re more thorough (and correct) with our analysis; we reinforce the system design; we add detail to the subsystems; and then we write some more code, code that stands more chance of meeting the requirements. Once our system is complete, perhaps after three or four spirals, we can perform rigorous testing and deploy the system. Compared to the waterfall methodology, we’re now working more like a sculptor creating a statue: we put together a basic framework of chicken-wire, then we add clay, layer by layer, until we achieve the desired effect. As we work: it becomes clearer and clearer how long our project will take; all our sponsors can see that progress is being made; and our conﬁdence that we can deliver a good result grows. We ﬁnish when everyone is happy. It looks as though the spiral methodology has fewer problems than the waterfall methodology: it involves our sponsors throughout the life cycle; everyone can see that we’re on track; it’s less rigid (we can tune the number of revolutions and how long each one takes). Altogether, it’s a better ﬁt to the creative nature of software development, as opposed to the engineering nature of building a bridge. The spiral methodology isn’t perfect though. The trouble is that we’ve simply attacked the waterfall by doing everything three or four times, which means that, although the problems have got smaller, they haven’t gone away altogether. We still have some inﬂexibility, because we’re supposed to proceed in an orderly manner through the classical phases; if we ﬁnd mistakes, we can’t ﬁx them until the next revolution. Therefore, the spiral methodology on its own is not much use – we need to combine it with something else.

5.4.2 Iterative Methodology So how do we improve the spiral methodology? To extend our sculpture analogy, the spiral methodology forces us to complete each layer of clay before we go on to the next. For

111

112

Chapter 5

Specification

Design

Analysis

Implementation

Testing

Deployment

Maintenance

Requirements

Figure 5.4: Iterative development methodology

example, let’s say that our sculpture is going to be of a person. Having put the ﬁrst layer of clay on the head, what if we decide that we want to work on the nose for a while before moving on to the rest of the body? What if, having added the torso, we decide that the forehead is now too narrow and warrants some immediate attention? What we need is a methodology that allows us to iterate over the phases, moving backwards and forwards, or round and round, as the need arises. This leads to the iterative methodology depicted in Figure 5.4. (Here, iterations are shown taking place within a spiral, but they could also be applied to a waterfall.) Now, we have a much more natural way of massaging our software from its early stages into a well-formed, elegant whole that satisﬁes all the sponsors. But, you’re probably wondering, how do we avoid chaos? We have at least three life-savers here: • The classical phases remind us what we should be doing at each stage and in which general direction we should be moving. • The artifacts (diagrams, descriptions, code, etc.) that we produce as we work within the classical phases do not get thrown away but are gradually improved as we move towards deployment. • The software production tools that support our chosen methodology and the notations help us to ensure that the artifacts are consistent and kept in one place. But we still have a problem. With the iterative and spiral methodologies, we’re still trying to eat the elephant: we’re still trying to deliver a complete system for deployment. So, we need to add a ﬁnal element: increments.

5.4.3 Incremental Methodology Let’s return to our sculpture analogy one last time. If we’re asked to produce a family of sculptures – the parents, the kids, the cat, the dog – it’s a good idea to deliver one sculpture

New Methodologies

at a time, so that our sponsors can see our progress as each piece is delivered; we get to sign off each piece of work and concentrate on the next; we can ask for incremental payment. In software development, this is the incremental methodology (see Figure 5.5). release

1.0

1.1

2.0

2.1

time

Figure 5.5: Incremental development

With the incremental methodology, we aim to deliver version 1.0 of our system with basic, critical functionality. Then, some time later, we deliver version 1.1 with additional functionality (as a replacement for version 1.0). Next, we might deliver version 2.0 with a whole raft of changes. And so on, throughout the lifetime of the system. Not only do we acknowledge from the start that we need several bites at the elephant, but we’re keeping up with changing requirements and a shifting marketplace. As you probably know from your own experience as a software purchaser, incremental delivery is what tends to happen in practice anyway, whether or not we plan for it. In other words, if we try to swallow the elephant whole, we fail. By planning for incremental delivery, we turn the perception of failure into a perception of wisdom. However, we must avoid at all costs the nightmare scenario of rewriting all our code for each new increment – this would be like an endless series of separate waterfalls. Thus, good analysis, good design, reusable code and extensibility become critical.

5.4.4 Combining the Methodologies So, the waterfall methodology is inadequate in most cases – although it does have the phases in the correct logical order – and the alternative methodologies (spiral, iterative and incremental) all have desirable properties but none of them is good enough on its own. So, we must combine all four in some way, but how? At the highest level, we know from the incremental methodology that we must plan a succession of increments. Within each increment, the spiral methodology suggests that we should have at least two attempts to produce each increment. Within each spiral, the waterfall methodology speciﬁes the phases and the order in which they occur. Within each mini-waterfall, the iterative methodology allows us to repeat phases, or combinations of phases, as we see ﬁt (for example, several cycles of requirements and analysis); the iterative methodology also allows us to ﬁx a problem as soon as we discover it (for example, we might

113

114

Chapter 5

discover during subsystem design that the system design makes some piece of functionality impossible, so we ﬁx the system design before we carry on). The combination of the methodologies is shown in Figure 5.6.

release

1.0

1.1

2.0

2.1

time

Figure 5.6: Combining spiral, iterative and incremental development

None of this discussion indicates how we should plan and schedule a particular project, which depends on the size of the project, the number of developers, the experience of the developers, the experience of the managers in planning and scheduling this kind of development, and so on. As far as this book is concerned, detailed planning and scheduling are management issues rather than software issues, so they’re not covered. Sufﬁce to say that, in object orientation, there should be no absolutes: it’s quite reasonable for the managers, mentors and experienced developers to decide on the increments, spirals, iterations and artifacts that are appropriate for each new development. Then, the planning must be adapted in the light of increasing knowledge and changing requirements. Although most theorists agree on the nature of the waterfall methodology, there is some disagreement about the terminology used for the others. In this book, you’ve seen particular deﬁnitions of the spiral, iterative and incremental methodologies but you should be aware that other people use these terms differently or even as synonyms. In particular, ‘spiral’ is sometimes referred to as ‘iterative’. However, the ideas discussed here are equally valid whatever name you give them.

5.5 OBJECT-ORIENTED METHODOLOGIES All object-orientation experts agree that a good methodology is essential for software development, especially when working in teams. Thus, quite a few methodologies have been invented over the last decade. Broadly speaking, all object-oriented methodologies are alike – they have similar phases and similar artifacts – but there are many small differences. Object-oriented methodologies tend not to be too prescriptive: the developers are given some choice about whether they use a particular type of diagram, for example. Therefore, the

Object-Oriented Methodologies

development team must select a methodology and agree which artifacts are to be produced, before they do any detailed planning or scheduling. In general, each methodology addresses: • • • • • •

The philosophy behind each of the phases. The workﬂows and the individual activities within each phase. The artifacts that should be produced (diagrams, textual descriptions and code). Dependencies between the artifacts. Notations for the different kinds of artifact. The need to model static structure and dynamic behavior.

Static modeling involves deciding what the logical or physical parts of the system should be and how they should be connected together. Dynamic modeling is about deciding how the static parts should collaborate. Roughly speaking, static modeling describes how we construct and initialize the system, while dynamic modeling describes how the system should behave when it’s running. Typically, we produce at least one static model and one dynamic model during each phase of the development. Some methodologies, especially the more comprehensive ones, have alternative development paths, geared to different types and sizes of development. The methodology used in this book, Ripple, is geared towards learning what’s involved in all software development, large or small, but it is applicable to the real world as well.

5.5.1 UML, RUP and XP By the mid-1990s, the best-known methodologies were those invented by Ivar Jacobson [Jacobson et al. 92], James Rumbaugh [Rumbaugh et al. 91] and Grady Booch [Booch 93]. Each had his own consulting company using his own methodology and his own notation. By 1996, Jacobson and Rumbaugh had joined Rational Corporation (founded by Booch), and they had developed a set of notations which became known as the Uniﬁed Modeling Language (UML) [OMG 03a]. The ‘three amigos’, as they have become known, donated UML to the Object Management Group (OMG) for safekeeping and improvement. OMG (www.omg.org) is a not-for-proﬁt industry consortium, founded in 1989 to promote open standards for enterprise-level object technology; their other well-known work is CORBA [OMG 04]. Some developers consider UML simply as a notation to be used for brainstorming and high-level documentation. Others consider UML to be a pictorial programming language, generating code from it or synthesizing it from existing code. UML was used as a notation during brainstorming of the case study for this book. At the end, the main diagrams were made to match the ﬁnished code exactly – thus, the end result looks the same as it would have done if UML had been used as a pictorial programming language.

115

116

Chapter 5

Once UML was in safe hands, the three amigos set about designing a methodology that drew on the best aspects of their individual work. Within a couple of years, they had come up with their own spiral, iterative and incremental method called the Rational Uniﬁed Process (RUP) [Jacobson et al. 99]. As you might expect, RUP is not the only methodology available, nor even the only one that uses UML as the notation. Another popular methodology is extreme programming (XP) [Beck 99]. XP is referred to as an agile methodology because it is responsive to change. XP is distinguished by two radical ideas: pair programming and test-driven development. With pair programming, all development is carried out by two people sitting in front of the screen rather than one – the idea is that, rather than improving the speed at which software is produced, pair programming improves the quality of the software (it also helps developers to accelerate their growth by the sharing of ideas). According to fans of test-driven development, continuous testing is so important that not only should it be done by the developers themselves, but the tests should be written before the code.

5.5.2 The Need for Development Tools To be effective, any spiral, iterative and incremental methodology requires an end-to-end development tool. The need to amend project artifacts iteratively is a strong justiﬁcation for the use of a software tool. Such a tool should allow members of the project team to produce the artifacts and then to store them. More speciﬁcally, it should support:

• Traceability: Recording the connections between an artifact and the artifacts derived from it, for example, recording which particular subsystem gave rise to a group of implemented classes. Most traceability information will be entered by developers rather than inferred by the tool. • Change history: Recording the changes that were made to artifacts, who made the changes and when. Where feasible (that is, with textual artifacts) the tool should be able to provide a summary of the differences between one edition of the artifact and another. • Multiuser access control: Making sure that simultaneous access to artifacts doesn’t cause problems. There are three mechanisms relevant here: authorization (controlling who can read the artifacts and who can edit them); multiuser read–single-user edit (only allowing one developer at a time to edit an artifact or group of artifacts, but allowing all authorized users to see the unedited version); versioning (allowing any number of developers to edit an artifact, each producing a distinct version of the artifact – at any time, only one version will be the ofﬁcial one). • Reduced redundancy: Ensuring that we never have to update anything in more than one place. Typically, information will appear in several artifacts at once. If we have a tool, it can treat the artifacts as alternative projections on a single model.

Ripple Overview

• Consistency checking: Ensuring that an artifact is consistent with related artifacts. It’s not always possible to enforce consistency. For example, in an ideal world, a tool would be able to check that the code written for a method conforms to its speciﬁcation, but modern computers simply can’t do that (it’s just too difﬁcult, or even impossible). Even in cases where the tool can enforce consistency (or traceability for that matter), the developer must be allowed to disable the check. For example, one developer might want to carry analysis classes through to the design phase, while another might prefer to start design with a new set of classes altogether: any tool that requires every analysis class to appear in the design model would be a help to the ﬁrst kind of developer and a hindrance to the second. • Networked operation: Providing access to all artifacts from any machine on the project network. These days, the best foundation for networked operation is TCP/IP, the basis of all protocols that operate over the Internet and individual intranets. • Testing the artifacts that we produce, as we go along. The most obvious case is where implementation code is tested for effectiveness and correctness, but the principle can be applied to other artifacts too (as in the case of recording the results of design review). Rational Corporation developed a tool called Rose, based on RUP and UML. Rose became probably the best-known object-oriented development tool. In 2003, Rational was bought by IBM and the tool has been reworked as a set of modular products (and is now called Rational Application Developer). Of course, there are many development tools available – Rational products are only mentioned here because of the historical perspective.

5.6 RIPPLE OVERVIEW In this book, you will ﬁnd object-oriented versions of all the classical phases of software development and you will see how they ﬁt into an object-oriented methodology. Because object orientation is so accessible, developers can be involved in all of the phases; customers can be involved in the early stages, which helps developers to do their job; and managers are not shut out of the developers’ world, so communication is improved. We have seen how the classical phases ﬁt into the ideal object-oriented methodology – one that draws on the best aspects of the spiral, iterative and incremental approaches. In subsequent chapters, we’ll look at each of the major phases that are carried out before the release of code: requirements, analysis, system design, subsystem design, speciﬁcation and testing. Implementation won’t be covered extensively, because that would require detailed knowledge of a particular programming language. Design patterns will be discussed, because they allow us to take implementation ideas off the shelf. By the end of this book, you will ﬁnd that you have traveled around one spiral of an initial increment (version 1.0, after an iterative spiral, if you like). Although this may look suspiciously like the waterfall methodology dressed up as something more elegant, it’s simply

117

118

Chapter 5

the nature of a book: content laid out end to end with no repetition. It can’t be emphasized enough that, when you come to try these techniques for yourself, you must be prepared to spiral, iterate and deliver incrementally. The case study used from here on, iCoot, was certainly not developed in a waterfall fashion. The artifacts included in Appendix B resulted from two increments, each comprising a number of iterative spirals. Since UML, the de facto standard notation, is used throughout this book, you can be sure that the diagrams that you see will be similar to the ones that you’ll encounter in the real world. For Ripple, UML notation is employed whenever possible. UML has 13 types of diagram. The UML speciﬁcation doesn’t say where these diagrams should be used in any particular methodology – we’re free to use whichever we think is appropriate at any stage. • Use case diagrams categorize the ways in which a system is used. • Class diagrams show classes and how they can be ﬁtted together (they can also show objects). • Object diagrams show only objects and how they can be ﬁtted together. • Activity diagrams show activity by humans or objects in a similar way to a ﬂow chart. • State machine diagrams show the various states of any object with an interesting or complicated life cycle. • Communication diagrams show the messages sent between objects in some scenario. • Sequence diagrams show similar information to communication diagrams, but emphasizing sequences rather than connections. • Package diagrams show how related classes are grouped together, for the beneﬁt of developers. • Deployment diagrams show machines, processes and deployed artifacts for a ﬁnished system. • Component diagrams show reusable components (objects or subsystems) and their interfaces. • Interaction overview diagrams show individual steps of an activity using sequence diagrams. • Timing diagrams show precise timing constraints for messages and object states. • Composite structure diagrams show how objects ﬁt together in an aggregation or composition, showing interfaces and collaborating objects. Table 5.1 summarizes the artifacts of Ripple, organized by phase. As you can see, some of these artifacts are found in UML and some are not. This is simply because UML doesn’t cover everything; to a large extent, it just allows us to draw pictures of our code. For the Ripple artifacts that are not covered by UML, alternative notation is used. Despite the fact that this notation is nonstandard, its content is based on widely accepted theory and practice.

Ripple Overview

Table 5.1: Ripple artifacts by phase Phase

Artifacts

UML

119

120

Chapter 5

Ripple is described progressively as we go through the chapters, but if you would like to read a quick summary at any point, go to Appendix 1. Since this book has an emphasis on software-related artifacts, the focus is on requirements, analysis, design and speciﬁcation. Glossaries, test plans and mission statements are also discussed. Other issues, such as management, implementation, deployment and maintenance won’t be covered in any detail. Although you will see an example of an activity diagram and a state machine diagram, they’re not used widely here and are generally considered optional. Also, not all of the UML diagram types are used. Component diagrams, interaction overview diagrams, timing diagrams and composite structure diagrams are simply not necessary for the purposes of this book. Most of these types of diagram can be expressed using the other diagrams, apart from timing diagrams, which are more useful for the design of real-time software and hardware. As well as the diagrams, UML has a class speciﬁcation language called Object Constraint Language (OCL). OCL won’t be covered, because it would require an entire book in its own right; however, in Chapter 12, you will ﬁnd a small example, to give you a taster. Most of the speciﬁcation discussion is informal, relating to comments in code and design artifacts. In the next few sections, you’ll see examples of some of the UML diagrams that are used for Ripple. You should bear in mind, though, that UML, like most comprehensive standards, is rather large. Therefore, for practical reasons, you will be shown only the fundamentals of the notation, with some of the ﬁner detail left out. Whenever you look at one of the diagrams in this book, bear in mind that UML allows us to suppress information that is not relevant to the discussion. For example, just because you only see labeled boxes when a class is being discussed, don’t assume that the class has no attributes or operations.

5.6.1 Use Case Diagram A use case is a static description of some way in which a system or a business is used, by its customers, its users or by other systems. A use case diagram shows how system use cases are related to each other and how the users can get at them. Each bubble on a use case diagram represents a use case and each stick person represents a user. Use case diagrams (static artifacts) are described in Chapter 6. Figure 5.7 depicts a car rental store accessible over the Internet. From this picture, we can extract a lot of information quite easily. For example, an Assistant can make a reservation; a Customer can look for car models; Members can log on; users must be logged on before they can make reservations; and so on. Each use case is more than just a title such as U7:Make Reservation or U13:Look for Car Models; it must include the actual steps involved in using the system or business. Although

Ripple Overview

iCoot >

<

U2: View Results

<<

<<extend>>

inc lud e

>

>>

nd

ex te

> d> ten x e <<

Customer

d>>

U7: Make Reservation

U5: Log On

<<

U10: View Reservations

> tend> <<ex > d> ten ex < <

<<exten

Member}

U6: View Member Details <<extend>

<<extend>>

a logged-On

NonMember >>

U4: Search

U3: View Car Model Details {Customer is

U13: Look for Car Models

U1: Browse Index

U12: Log Off

U11: Cancel Reservation

U9: Change Password

U8: View Rentals

Member

Assistant

Figure 5.7: A use case diagram

UML speciﬁes a notation for use case diagrams, it doesn’t do so for the steps of the use case itself. RUP refers to the steps of a use case, and a few other bits and pieces, as use case details. The details for U3:View Car Model Details are shown in Figure 5.8. It should be easy for you to appreciate that viewing car model details involves a customer selecting a car model, requesting its details, and then receiving speciﬁc information about the car model in return. Use case details (dynamic artifacts) are described in Chapter 6.

5.6.2 Class Diagram (Analysis Level) A class diagram shows which classes exist in the business (during analysis) or in the system itself (during subsystem design). Figure 5.9 shows an example of an analysis-level class diagram, with each class represented as a labeled box. Class diagrams (static artifacts) are introduced in Chapter 7. As well as the classes themselves, a class diagram shows how objects of these classes can be connected together. For example, Figure 5.9 shows that a CarModel has inside it a CarModelDetails, referred to as its details.

121

122

Chapter 5

U3: View Car Model Details. (Extends U2, extended by U7.) Preconditions: None. 1. Customer selects one of the matching Car Models. 2. Customer requests details of the selected Car Model. 3. iCoot displays details of the selected Car Model (make, engine size, price, description, advert and poster). 4. If Customer is a logged-on Member, extend with U7. Postconditions: iCoot has displayed details of selected Car Models. NonFunctional Requirements: r1. Adverts should be displayed using a streaming protocol rather than requiring a download.

Figure 5.8: Details of a system use case

1..*

Car details

*

Rental

taken out by

Car Details

1 example of

Make

*

0..1

*

1

1..

rented under

*

Car Model Details

Customer

1 1

makes

*

details

Car Model

*

classified as 1

**

sold by

*

1..

*

1..

Reservation

guaranteed by

logs in with

lives at

1 1

Category

NonMember

1

Member

Vendor

Credit Card

1

Internet Account

1 Address

Figure 5.9: A class diagram at the analysis level

5.6.3 Communication Diagram A communication diagram, as its name suggests, shows collaborations between objects. The one shown in Figure 5.10 describes the process of reserving a car model over the Internet: A Member tells the MemberUI to reserve a CarModel; the MemberUI tells the ReservationHome to

Ripple Overview

m:Member c:CarModel

1:reserve(c) 1.1:fineWarning() 1.1.1:confirmReserve() 1.1.1.3:reservationNumber(n)

Member

r()

um

be

N et

=g

:MemberUI

1.

1.

1.

:n

2 1.

. .1

1

1:

r=

cr

ea

te

(c ,m

)

r:Reservation

:Reservation Home

Figure 5.10: A communication diagram

create a Reservation for the given CarModel and the current Member; the MemberUI then asks the new Reservation for its number and returns this to the Member. Communication diagrams (dynamic artifacts) are described in Chapter 7.

5.6.4 Deployment Diagram A deployment diagram (see Figure 5.11) shows how the ﬁnished system will be deployed on one or more machines. A deployment diagram can include all sorts of features such as machines, processes, ﬁles and dependencies. Figure 5.11 shows that any number of HTMLClient nodes (each hosting a WebBrowser) and GUIClient nodes communicate with two server machines, each hosting a WebServer and

123

124

Chapter 5

*

*

<<device>> CootHTMLClient <<device>> CootGUI Client

<<ExecutionEnvironment>> WebBrowser

JRMP

HTTP

2

<<device>> CootServer

<<ExecutionEnvironment>> WebServer

<<ExecutionEnvironment>> CootBusinessServer

2

<<device>> DBServer

<<ExecutionEnvironment>> DBMS

Figure 5.11: A deployment diagram

a CootBusinessServer; each WebServer communicates with a CootBusinessServer; and each CootBusinessServer communicates with a DBMS running on one of two DBServer nodes. Deployment diagrams (static artifacts) are described in Chapter 8.

5.6.5 Class Diagram (Design Level) The class diagram shown in Figure 5.12 uses the same notation as the one introduced in Figure 5.9. The only difference is that design-level class diagrams tend to use more of the available notation, because they’re more detailed. This one expands on part of the analysis class diagram to show methods, constructors and navigability. Design-level class diagrams (static artifacts) are described in Chapter 10.

Ripple Overview

-address 1

1

Store

Address

~Store(:Address) <> +getAddress():Address +addCar(:Car)

~Address(:String,:String,:String,:String) <> +getHouse():String +getCounty():String +getPostcode():String

1

*

-cars:Set CarModel

Car ~CarModel(:Category,:Make,:String,:int) <> +getCategory(): Category +getMake():Make +getModelNumber():String +getPrice():int

~Car(:CarModel,:int) <> +getModel():CarModel +getTraveled():int +setTraveled(:int) *

1

1

*

-model 1 Make

-makes:List 1..*

~Make(:String) <> getName():String

-category

Category ~Category(:String) <> +getName():String

Figure 5.12: A design-level class diagram

5.6.6 Sequence Diagram A sequence diagram shows interactions between objects. Communication diagrams also show interactions between objects, but in a way that emphasizes links rather than sequence. In this book, sequence diagrams are used during subsystem design, but they’re equally applicable to dynamic modeling during analysis, system design and even requirements capture. The diagram in Figure 5.13 speciﬁes how a Member can log off from the system. Messages are shown as arrows ﬂowing between vertical bars that represent objects (each object is named at the top of its bar). Time ﬂows down the page on a sequence diagram. So, Figure 5.13 speciﬁes, in brief: a Member asks the AuthenticationServlet to logoff; the AuthenticationServlet passes the request on

125

126

Chapter 5

sd U12

Member

:Authentication Servlet

:Authentication Server

:Member Home

m:Member

:Internet Account

logoff() logoff(id) m=findBySessionId(id) id retrieved from browser session

setSessionId(0) setSessionId(0)

home page

Figure 5.13: A sequence diagram from the design phase

to the AuthenticationServer, reading the id from the browser session; the AuthenticationServer ﬁnds the corresponding Member object and tells it to set its session id to 0; the Member passes this request on to its InternetAccount; ﬁnally, the Member is presented with the home page. Sequence diagrams (dynamic artifacts) are described in detail in Chapter 10.

5.7 SUMMARY In this chapter, we looked at: • The classical phases of software production – requirements, analysis, system design, subsystem design, speciﬁcation, implementation, testing, deployment and maintenance – and how they can be used in a combination of spiral, iterative and incremental methodologies. • Static modeling, which describes how we construct a system, and dynamic modeling, which describes how the system should behave when it’s running. • The UML notation and the Ripple methodology that is used in the rest of this book.

FURTHER READING Steve McConnell provides a good overview of managing a successful software project [McConnell 98], from planning through to testing. It is an easy read because of its high-level approach and conversational style.

Review Questions

The original book on Rational’s Uniﬁed Process [Jacobson et al. 99] provides a good overview of the theory underlying RUP. For the current status of RUP, take a look at the Rational web site, www.rational.com. For more information on XP, see [Beck 99] by one of its inventors or the following web sites: www.extremeprogramming.org and www.xprogramming.com. To ﬁnd out more about agile development in general, see [Cockburn 01] and the Manifesto of Agile Software Development at www.agileManifesto.org. The best-known introduction to UML is [Fowler 03]. Although Martin Fowler doesn’t cover the whole of UML, his book is a good stepping-stone between the book you’re currently reading and the UML Speciﬁcation itself [OMG 03a]. For the serious UML user, the speciﬁcation is still an essential tool for answering questions about obscure pieces of notation and the semantics of the language itself, so that you can ensure that your diagrams are correct. Speciﬁcations are hard to read at the best of times; with the UML Speciﬁcation it can be a shock to ﬁnd that UML is deﬁned using UML, but it turns out to be a good idea: it means that the descriptions are precise; it is a good opportunity to see how UML can be used to document a large and complicated object-oriented model; and it allows you to see what is and is not legal according to the model.

REVIEW QUESTIONS 1. Which of the following UML artifacts are used to show the distribution of processes, resources and objects in a system? Choose only one option. (a) (b) (c) (d) (e) (f) (g)

Interaction diagrams. Sequence diagrams. Deployment diagrams. Communication diagrams. State machine diagrams. Class diagrams. Glossaries.

2. What are the traditional steps in software production? Choose all options that apply. (a) (b) (c) (d) (e) (f) (g)

Maintenance. Design. Iteration. Incrementation. Deployment. Analysis. Requirements Capture.

127

128

Chapter 5

(h) (i) (j) (k)

Testing. Reuse. Implementation. Speciﬁcation.

ANSWERS TO REVIEW QUESTIONS 1. The UML artifacts used to show the distribution of processes, resources and objects in a system are c. Deployment diagrams. 2. The traditional steps in software production are: (a) Maintenance. (b) Design. (e) Deployment. (f) Analysis. (g) Requirements Capture. (h) Testing. (j) Implementation. (k) Speciﬁcation.

II

Understanding the Problem

6

Gathering Requirements

I

n this chapter we’ll look at the requirements phase of software development, detailing the starting points and introducing use cases. We also look at how to model the business into which our system will ﬁt by identifying and validating business use cases.

Learning Objectives Understand the purpose of the requirements phase.

•

Model the business context and system functionality.

Record the system requirements on a complete use case model.

Chapter Outline 6.1 Introduction 6.2 The Birth of a System 6.3 Use Cases 6.4 Business Perspective 6.5 Developer Perspective 6.6 Summary Further Reading Review Questions Answers to Review Questions

6

132

Chapter 6

6.1 INTRODUCTION The aim of the requirements phase is twofold: • Examine the business context: We need to clarify the reasons for wanting the software to be developed in the ﬁrst place – if we can’t come up with good reasons, we shouldn’t write the software at all. When we’ve decided that we do want to produce a software system, we need to make sure that we understand the business and that our understanding matches that of our sponsors – this is also a good opportunity to clarify just who the sponsors are. • Describe the system requirements: This involves not only deciding on the functionality of the system but also detecting any constraints – performance, development cost, resources and so on. We would expect system requirements to form part of the requirements phase, but why do we model the business? Figure 6.1 shows the alternative to a well-deﬁned requirements phase. These two developers start with some vague idea of the kind of system that they believe ought to be produced, while paying only grudging attention to their sponsors. This blinkered approach is common among novice programmers who don’t yet know what they’re doing, but we would hope that little of this attitude remains by the time we have become serious professionals.

I'll start coding, you go and see what the customers want

Figure 6.1: Self-taught developers

The Birth of a System

The tendency to dive into coding is not only born of arrogance. It can come from fear: ‘We’re not sure we can produce what the sponsors need, but we know what we can produce: hopefully, once we’ve ﬁnished, we can persuade everyone that what we have is what they really needed in the ﬁrst place’. Difﬁcult as it may be to stay away from our keyboards, we must ﬁrst make sure that we understand the business context for the new system, then work with our sponsors to agree what the system will do. The term ‘sponsor’ is used to mean anyone who has an interest in seeing the ﬁnal system delivered: for example, the internal or external customer commissioning the system, the potential end users, the managers, and even the shareholders. Before we even consider writing a piece of software, we must investigate the business in which the software will operate – without a thorough understanding of the business, we could hardly expect to produce something that would enhance that business. (We must be aiming to enhance what’s already there, otherwise why would we go to the trouble of writing software in the ﬁrst place?) The term business is used here in its loosest sense: although, admittedly, this book has a bias towards business systems in the sense of banking, administration, e-commerce and the like, most of what is presented applies equally well to scientiﬁc systems, home systems, or anything else that requires software. If you prefer, think of ‘business’ as ‘problem domain’. Once we have understood the business and documented our understanding as business requirements, we need to think about what our software will do for the users. Deciding what the system needs to be able to do and, just as importantly, what it should not do, will help us to focus on producing only the necessary code. Without a thorough understanding of the system requirements, we would risk wasting time on developing code that we’re not being paid to produce. (Again, ‘being paid’ shouldn’t be taken too literally – you may be writing a small system purely for your own beneﬁt, in which case you’ll be more concerned with not wasting your time.) System requirements are commonly separated into two categories: functional and nonfunctional. Functional requirements are the things the system must be able to do, i.e. the services that it must provide in response to external stimuli, such as ‘browsing the catalog’ and ‘reserving a car model’. Nonfunctional requirements are everything else that needs to be speciﬁed. Nonfunctional requirements might include the client Web browsers that must be supported, the use of streaming video (as opposed to downloadable ﬁles) for adverts, a user interface that can be used easily by novice Internet surfers, and so on.

6.2 THE BIRTH OF A SYSTEM Every system starts somewhere. We may be lucky enough to get a detailed document from the customer, usually with proprietary layout and content. Or, we may simply be presented

133

134

Chapter 6

with something like a mission statement, a short statement of some new, desirable, business direction. As developers, we must transform the customer’s requirements document or mission statement into a complete, unambiguous description of the system to be developed, in a standard format that the customer can understand and ratify. Admittedly, ‘complete’ and ‘unambiguous’ are impossible to achieve in practice. We shouldn’t expect to come close to these goals the ﬁrst time round. However, it’s still useful to know that eventually we will have a document that describes everything the system will do (and, by omission, everything that it won’t do), with little room for misinterpretation.

Case Study Nowhere Cars mission statement Since we automated the tracking of cars at our stores – using bar codes, countertop terminals and laser readers – we have seen many beneﬁts: the productivity of our rental assistants has increased 20%, cars rarely go missing and our customer base has grown strongly (according to our market research, this is at least partly due to the improved perception of professionalism and efﬁciency). The management feels that the Internet offers further exciting opportunities for increasing efﬁciency and reducing costs. For example, rather than printing catalogs of available cars, we could make the catalog available to every Internet surfer for browsing on-line. For privileged customers, we could provide extra services, such as reservations, at the click of a button. Our target saving in this area is a reduction of 15% in the cost of running each store. Within two years, using the full power of e-commerce, we aim to offer all of our services via a Web browser, with delivery and pick-up at the customer’s home, thus achieving our ultimate goal of the virtual rental company, with minimal running costs relative to walk-in stores. Even this three-paragraph mission statement contains a lot of information: history of automation at the company; customer satisfaction to date; on-line catalogs and reservations; privileged and non-privileged customers; savings history and savings targets; company end-game (the ‘virtual rental company’). Admittedly, some of the management’s dreams are a long way off (it may be more than two years before customers are comfortable with virtual rental stores), but at least we have two good starting points for our investigation: What services do the company’s stores currently offer? Which of those services are appropriate for Internet delivery?

Use Cases

Case Study (cont’d) The mission statement above is the basis for the case study used in the rest of this book. The ﬁctitious company’s new system is referred to as Coot, with the Internet facilities available to customers referred to as iCoot. The unique selling point of Nowhere Cars is that they rent specialist cars to wealthy enthusiasts for extended periods. Since the supply of each kind of car is limited, customers must turn up at a store when they actually want to rent. Cars are rented on a ﬁrst-come, ﬁrst-served basis and customers can take their pick from what is currently available. Alternatively, customers who are keen to rent a model of car which is not available can make a reservation. An assistant will contact the customer directly when a matching car becomes available; the customer must collect it within two days (or pay a levy for depriving other customers of the car). As yet, there are no home delivery or home pick-up services (partly for insurance reasons). For members, who must register, reservations can be made by telephone.

6.3 USE CASES Ivar Jacobson invented use cases to deﬁne the way in which part of a business or a system is used [Jacobson et al. 92]. Although, at ﬁrst sight, use cases appear more process-oriented than object-oriented, they’re widely considered to be the most effective tool for describing a system’s functional requirements. Most nonfunctional requirements can be recorded alongside a closely-related use case (any others can be listed separately). The use cases in this book contain all the essential elements in a format which is convenient for learning. In this book, use cases are used to document our understanding of the way a business operates – business requirements modeling – and to specify what our new software system should be able to do – system requirements modeling. The business use cases in this book use an informal, descriptive style: they describe, for the beneﬁt of nonexperts, something that already exists. The system use cases, on the other hand, will be more prescriptive: they specify, mainly for the beneﬁt of software developers, exactly what functionality needs to be implemented. A use case starts with a participant called an actor; it then descends into the business, or the system, and eventually returns to the actor. The effect of each use case should be of value to the actor (otherwise, why would they initiate it in the ﬁrst place?). Of course, value can mean different things to different people: it could be some information that the actor wishes to retrieve, some effect that the actor wishes to have on the system, some money, a purchase, or pretty much anything else that might motivate them. Being driven by use cases,

135

136

Chapter 6

far from sending us down the traditional path, actually helps us to ﬁnd objects, attributes and operations. Case Study Nowhere Cars use cases • Member Reserves Car Model is a business use case that describes how, according to current practice, a member makes a reservation. It may be couched in terms that would apply to any car-rental business or it may bring in details speciﬁc to the way that Nowhere Cars operates. We look for business use cases during business modeling, the ﬁrst step of requirements capture. A business use case may refer to existing software systems, or it may not involve computers at all – for example, the last time you telephoned a car-rental shop and asked them to reserve a particular car model, were you able to tell if the assistant at the other end of the line, in the course of performing the transaction, used a computer or a pad and paper? More to the point, did you care? • Make Reservation is a system use case that describes how the system that we intend to produce will allow Nowhere Cars to conduct reservations over the Internet. A system use case describes a service that the new, or replacement, system will provide: in this example, the member deﬁnitely does use some software – a Web browser and a back end server. Part of our job is to specify exactly what input the user should provide and what they should expect to get in return.

For the sake of simplicity, use cases, especially system use cases, should not overlap. A use case is written in natural language, broken up into a sequence of steps. Diagrams can accompany the use case if more explanation is needed.

6.4 BUSINESS PERSPECTIVE In this section, we’ll see how to put together a model of the business, as a precursor to modeling the functionality of the proposed system. A business model can be as simple as a class diagram, showing the relationships between business entities – this is sometimes referred to as a domain model. A domain model may be sufﬁcient for small projects, however, for most projects, we would want to produce an entire business model representing how the business operates, or at least that part of the business that surrounds the system we expect to develop. Use cases are not the only way of modeling a business, but they’re simple. More complex alternatives include business process modeling and workﬂow analysis. Use cases are simple because producing one doesn’t require specialist knowledge, just common sense and a

Business Perspective

certain amount of logic. The use case model that we produce here will contain the use cases themselves plus some other bits and pieces: • • • • •

Actor list (with descriptions). Glossary. Use cases (with descriptions and details). (Optional) Communication diagrams. (Optional) Activity diagrams.

UML deﬁnes the notation and semantics for activity diagrams and communication diagrams. The other artifacts are recommended by Jacobson. We’ll look at the business model components one by one, in the order in which you would typically create them in real life. However, bear in mind at all times that this is not a rigid workﬂow: as with any aspect of object-oriented development, we can iterate forwards, backwards, or round and round, until we have a complete picture.

6.4.1 Identifying Business Actors The ﬁrst thing we need to do is to identify the business actors. An actor is either a person playing some role within the business (as you might expect from the name), a department, or a separate software system. The reason for identifying departments and systems as actors is that, in logical terms, they interact as if they were people themselves: we’re interested in who (or what) initiates interactions and the sequence of steps. We don’t care whether a particular actor is ‘implemented’ as a person, a department or a piece of software. Identifying actors helps us to identify the ways in which the business is used, which will, in turn, indicate what the use cases are. Just as in real life, an actor can play different business roles at different times. For example, Fred Bloggs might be an assistant within the Nowhere Cars store until he clocks off; if he decides to rent a car before going home, he becomes a customer. At this stage of the development, you will be working with the other sponsors (principally the customers) to ﬁnd out how the business operates – the actors should fall out of your discussions easily.

Case Study Nowhere Cars business actor list • Assistant: An employee at one of our stores who helps Customers to rent Cars and reserve Car Models. • Customer: A person who pays us money in return for one of our standard services.

137

138

Chapter 6

Case Study (cont’d) • Member: A Customer whose identity and credit-worthiness have been validated and who, therefore, has access to special services (such as making Reservations by phone or over the Internet). • NonMember: A Customer whose identity and credit-worthiness have not been checked and who, therefore, must provide a deposit to make a Reservation or surrender a copy of their license to rent a Car. • Auk: The pre-existing system that handles Customer details, Reservations, Rentals and the catalog of available Car Models. • DebtDepartment: The department of Nowhere Cars that deals with unpaid fees. • LegalDepartment: The department of Nowhere Cars that deals with accidents in which a rented Car has been involved.

6.4.2 Writing the Project Glossary Even at this early stage, it’s a good idea to start maintaining a glossary – the modern alternative to a data dictionary. The phrase ‘data dictionary’ includes the word ‘data’, the kind of thing that object-oriented theoreticians are uncomfortable with, because it implies that data is being modeled in isolation. Separating data from processes was the old way of doing things: it’s much better to keep data and process together, hence the less emotive term ‘glossary’. A glossary de-mystiﬁes the jargon for anyone examining our software development artifacts. It also allows us to ﬁle away groups of synonyms, leaving us free to use one of each throughout the rest of the documentation. Case Study Nowhere Cars Glossary Term (Business object) (Business object) Customer (Business actor, business object) Member (Business object)

Deﬁnition

CarModel

Instance of a CarModel kept by a Store for Rental purposes. A model in our Catalog, available for Reservation.

...

A person who pays us money in return for one of our standard services. A Customer whose identity and credit-worthiness have been validated and who, therefore, has access to special services (such as making Reservations by phone, or over the Internet).

Car

Business Perspective

Each entry in the glossary deﬁnes a term – the deﬁnition can be short or long, as appropriate. The actor descriptions that we’ve seen so far are good starting points as glossary deﬁnitions, but the glossary deﬁnitions will often end up being more general, because most of the terms will apply in several contexts. As you can see from the case study glossary, you can record the relationships that each term has to the development phases (business actor, system actor, and so on). Below is a list of relationships that you can use (each entry may qualify as more than one of these): • • • • • • • • • •

Business actor: An actor appearing in the business requirements. Business object: An object appearing in the business requirements. System actor: An actor appearing in the system requirements. System object: An object appearing (inside the system) in the system requirements. Analysis object: An object appearing in the analysis model. Deployment artifact: Something deployed in the system, such as a ﬁle. Design object: An object appearing in the design model. Design node: A computer or process that forms part of the system architecture. Design layer: A vertical partition of a subsystem. Design package: A logical grouping of classes, used to organize the development.

In each of these cases, object means entity or ‘encapsulated data and process’, as usual. Each category of object – business, system, analysis or design – is subtly different, with some objects qualifying in more than one category. For example, when a Customer rents a Car, we’re dealing with a business object that is external to the system – the physical vehicle in the display area – and a system object that is inside the system itself, instantiated from a class that we implement. Glossary entries use the class naming style (words run together with initial capitals). As long as we use the same style in all our project documentation, it will be obvious to the reader that a deﬁnition exists in the glossary.

6.4.3 Identifying Business Use Cases Once we have the actors, the next task is to identify the business use cases. Each use case is a snippet of the business. At this stage, use cases may involve two-way communication between a number of actors, especially if they’re human actors. Later on, we’ll see that system use cases are more structured, because people normally tell the system what to do, rather than the other way round. There’s no set rule for deciding how to break the business down into use cases – common sense, logic and experience will help, as usual. Working with the sponsors (which you should be doing anyway) will also help. While talking to an assistant on the shop ﬂoor, for example,

139

140

Chapter 6

try to identify the different tasks that they ﬁnd themselves doing every day. Since assistants interact with customers, they should also be able to identify the ways in which customers use the business. Employee and management training manuals, mission statements, proprietary requirements documents, sales brochures and other kinds of document can also provide inspiration. At all times, when trying to ﬁnd use cases, keep the following question at the back of your mind: ‘What are the key activities that make this business work?’

Case Study iCoot Business use case list • B1:Customer Rents Car: Customer rents a Car that they have selected from those available. • B2:Member Reserves CarModel: Member asks to be notiﬁed when a CarModel becomes available. • B3:NonMember Reserves CarModel: NonMember pays a deposit to be notiﬁed when a CarModel becomes available • B4:Customer Cancels Reservation: Customer cancels an unconcluded Reservation, by phone or in person. • B5:Customer Returns Car: Customer returns a Car that they have rented. • B6:Customer Told CarModel is Available: Customer is contacted by an Assistant when a Car becomes available. • B7:Car Reported Missing: Customer or Assistant discovers that a Car is missing. • B8:Customer Renews Reservation: Customer renews a Reservation that has been outstanding for more than a week. • B9:Customer Accesses Catalog: Customers browse the catalog, in-store or at home. • B10: Customer Fined for Uncollected Reservation: Customer fails to collect a Car that they have reserved. • B11:Customer Collects Reserved Car: Customer collects a Car that they have reserved. • B12:Customer Becomes a Member: Customer provides CreditCard details and proof of Address to become a Member. • B13:Customer Notiﬁed Car is Overdue: Assistant contacts Customer to warn them that a Car they have rented is more than a week overdue. • B14:Customer Loses Keys: Replacement Keys are provided for a Customer who has lost them. • B15:MembershipCard is Renewed: Assistant contacts Member to renew membership when their CreditCard has expired. • B16:Car is Unreturnable: A Car is wrecked or breaks down.

Business Perspective

Case Study (cont’d) We have now started to narrow down the business to the area we’re particularly interested in. We know from the original mission statement that our customer is only hoping to make parts of the rental and reservation business available over the Internet so, for example, the sale of ex-rental cars has not been modeled.

B3: NonMember Reserves CarModel. 1. NonMember tells Assistant which CarModel to reserve. 2. Assistant finds CarModel on Auk. 3. 4. 5. 6.

Assistant asks for a deposit for the Reservation. Assistant asks for NonMember’s License and phone number. Assistant checks License visually. If License looks okay, assistant creates new Reservation and

records License number, phone number and a scan of the License in Auk. 7. Assistant gives NonMember a ReservationSlip containing the unique reservation number.

Figure 6.2: A business use case for Nowhere Cars

Remember that, during business modeling, we’re not interested in the way that our new system might operate. At this stage, we’re simply trying to describe the way the business currently operates. This may, or may not, involve existing software. The numbering scheme is arbitrary: UML does not specify a scheme and the list does not imply an order. Once we have a list of candidate use cases, we can list the steps involved in each one. UML speciﬁes nothing about the contents of a use case (or about the numbering or the descriptions, for that matter). Thus, we’re free to use natural language, step-by-step descriptions, structured language (natural language with if–then–else and loop structures), or whatever. Using steps rather than natural language encourages us to restrict ourselves to the bare bones of a use case. If we were to use structured language, we would be in danger of making our descriptions too algorithmic (computer-oriented). To make the use cases clear and independent of the implementation, unstructured steps will be used for the use case details in this book (see Figure 6.2).

6.4.4 Illustrating Use Cases on a Communication Diagram As well as writing down use case details, we can provide an illustration of use cases using communication diagrams. A communication diagram shows a series of interactions between

141

142

Chapter 6

actors and objects. A sequence diagram focuses on the interactions themselves and the order in which they take place. In order to discourage the use of too much technical detail early in the development process, the communication diagram is preferred for business modeling. Because UML is designed to apply to every possible situation, there is a lot of notation available to the developer for use in each of the different kinds of diagram. In this book, so as to avoid covering every possible detail, only the essential parts are used. Alternative ways of illustrating the same piece of information and anything too complex will also be ignored. Figure 6.3 shows ﬁve communicating elements. The nature of each element is shown by the icon used to represent it: a stick person represents an actor; a circle standing on a line represents a business object or entity; a circle connected to a vertical bar represents a boundary (something that manages the interaction between other elements – usually this is a piece of software, but it could also be a person); a stick person inside a boundary icon represents a human actor playing some kind of interface role (this shows that the boundary is a person rather than a piece of software).

2:find CarModel

1:reserve CarModel

Non Member

3:get deposit, License 5:confirm Reservation

Assistant

4:create Reservation

CarModel

Auk Interface

Reservation

Figure 6.3: A communication diagram for B3:NonMember Reserves CarModel

Without too much knowledge of the notation, it’s easy to appreciate from Figure 6.3 that making a reservation involves a nonmember, an assistant acting as a business boundary, a piece of software called Auk (acting as a boundary to some system) and two business objects. Since, at this stage, we’re modeling an existing way of doing things, we know that the interface and the system that it accesses must already be implemented and deployed. There is no mention of any software that we might eventually produce ourselves, because we haven’t got to that stage yet. A line connecting two elements on a communication diagram indicates that the elements can interact. So, we can see from Figure 6.3 that: the NonMember is requesting services from the Assistant; the Assistant is requesting services from the AukInterface; the AukInterface is requesting services from a CarModel and a Reservation. Since the CarModel and Reservation

Business Perspective

are accessed by a software interface, the implication is that they’re software objects inside the existing system, rather than physical objects outside (the distinction may or may not be important to us). As well as icons and connections, Figure 6.3 depicts individual interactions as numbered labels with an associated arrow. You can think of these as messages being sent from one element to another: the number indicates the message’s position in the sequence. So, we can interpret the entire collaboration as: • The NonMember asks the Assistant to reserve a CarModel. • The Assistant asks the AukInterface to ﬁnd the CarModel (this involves the CarModel object in some unspeciﬁed way). • The Assistant asks the NonMember for a deposit and a License. • The Assistant asks the AukInterface to create a Reservation (this involves a Reservation object in some way). • The Assistant conﬁrms the Reservation to the NonMember. You might think that this sequence of interactions should exactly match the sequence of steps in the use case details. However, because natural language is not a sequence of steps, a one-to-one match is unlikely. It’s more likely that each interaction will represent a summary of one or more steps. Despite the lack of an exact match with a use case, a communication diagram is still useful because it enhances the use case details and can help us to produce the details in the ﬁrst place. Since the interactions we deal with at this early stage are straightforward, it’s reasonable for us to produce one communication diagram per use case, but no more. For brevity, any collaboration that we depict should be the normal path through the use case. When we deal with system use cases, we can be more speciﬁc about abnormal paths, but for now we should be able to imply their existence within the use case itself: for example, step 6 in our business use case (Figure 6.2) begins If License looks okay: this implies that sometimes the License is invalid, but what the Assistant should do in that case has not been speciﬁed.

6.4.5 Illustrating Use Cases on an Activity Diagram UML includes another kind of diagram that can be useful during business modeling. An activity diagram shows dependencies between (parallel) activities as we move from an initial starting point to a desired goal. They are similar to ﬂow charts or Petri nets, traditionally used to model program ﬂow or human activity. Figure 6.4 shows an activity diagram being used to illustrate a business use case. Each rounded box in an activity diagram represents an action; an open ended arrow (an edge) indicates that the source action must be completed before the destination action is

143

144

Chapter 6

(NonMember) Tell Assistant which CarModel

(Assistant) Ask for Deposit and License

(NonMember) Find deposit

(NonMember) Find License

(Assistant) Find CarModel on Auk

(Assistant) Check deposit and License

[not ok] [ok] (Assistant) Make Reservation

Figure 6.4: An activity diagram for B3:NonMember Reserves CarModel

started; a black dot indicates the starting point for the activity; a black dot inside a white circle indicates the end of the activity; a diamond represents a decision – a guard on a departing edge indicates the reason for following that edge; thick black lines, known as forks and joins, are used to indicate the beginning and end of a set of concurrent actions. For each action, we can show who or what is responsible for the action by putting a name in parentheses, before the name of the action itself. The name identiﬁes a partition within the activity, and can be used to identify actors, departments, systems or objects. Partitions can also be indicated by grouping actions into rows, columns or cells.

Developer Perspective

We can see quite easily from Figure 6.4 that, to reserve a CarModel: 1. The NonMember tells the Assistant which CarModel to reserve. 2. The Assistant asks for a deposit and License. 3. While the NonMember is looking for the deposit and License, the Assistant looks for the CarModel on the Auk system. 4. Once everything has been found, the Assistant checks the deposit and the License. 5. If the deposit and License are valid, the Assistant makes the Reservation and the activity is ﬁnished. 6. Otherwise, the activity is ﬁnished. As with communication diagrams, we should not expect this interpretation to match the use case details step-for-step. In common with many UML diagrams, activity diagrams can be used for more than one purpose. For example, an activity diagram can be used to construct an entire business model or to document the algorithm employed by some software object.

6.5 DEVELOPER PERSPECTIVE The second part of requirements capture is modeling the software that we’re going to develop in order to improve the business. Regardless of whether you choose to document the business using a simple domain model, a fully-ﬂedged use case model, or something even more detailed such as a business process model or workﬂow analysis, it is widely accepted that the requirements of the software system should be captured using a use case model. This is because use cases are relatively easy to produce and easy for all sponsors to understand. The use case model for a system is more detailed and more prescriptive than for a business. For Ripple, the system use case model comprises: • • • • • •

An actor list (with descriptions). A use case list (with descriptions). A use case diagram. Use case details (including any related nonfunctional requirements). A use case survey. Supplementary requirements (system requirements that don’t ﬁt with any particular use case). • User interface sketches. • An enhanced glossary. • Use case priorities. We’ll look at how to produce each of these artifacts in turn.

145

146

Chapter 6

The list above includes several artifacts that we haven’t come across before. Even the ones that we’ve already seen during business modeling generally include more detail in system modeling. Communication diagrams have not been included; although there is nothing to stop you using them to illustrate your system use cases at this stage, for Ripple they have been deferred to a later phase in the development (dynamic analysis), where they’re considered more important. We should also spend some time evaluating existing systems. Such a system is referred to as a legacy system, because we inherited it as part of the existing business.

Case Study Nowhere Cars legacy system We need to decide whether Auk can reasonably be extended, or whether we should replace it altogether. This decision is not an easy one to make: on the one hand, we have a ﬁnished system that has been up and running for some time, one that assistants are familiar with; on the other hand, it may be too difﬁcult to open up Auk in order to add new facilities, or to write new software that communicates with it process-to-process (which may be inefﬁcient). Let’s assume that we have made the decision to replace Auk with a whole new system that will be compatible with Internet access and will support the ‘virtual rental store’ that our customer desires. Our new system will be called Coot. To alleviate the problems of staff training, the new interfaces (counter-top terminals and laser readers) will look and feel similar to the Auk ones. For the purposes of this book, we don’t need to examine all the facilities of Coot. Instead, we can concentrate on those parts of the system that provide the ﬁrst round of Internet facilities to customers. This cut-down set of facilities is called iCoot.

Identifying System Actors The ﬁrst thing we need to do is to identify and describe the system actors, with the help of our sponsors. The actors we identify at this stage should include only the people (and external systems) that interact directly with our proposed system, rather than actors from the wider business context. Case Study iCoot system actor list • Customer: A person using a Web browser to access iCoot.

Developer Perspective

Case Study (cont’d) • Member: A Customer who has presented their name, address and CreditCard details at one of our Stores; each Member is given an Internet password to accompany their membership number. • Assistant: An employee at a Store who contacts Members to tell them about the progress of their Reservations.

Identifying System Use Cases Once we have actors, we can look for use cases, again asking our sponsors for help. Each use case must have a short description.

Case Study iCoot system use case list • U1:Browse Index: A Customer browses the index of CarModels. • U2:View Results: A Customer is shown the subset of CarModels that was retrieved. • U3:View CarModel Details: A Customer is shown the details of a retrieved CarModel, such as description and advert. • U4:Search: A Customer searches for CarModels by specifying Categories, Makes and engine sizes. • U5:Log On: A Member logs on to iCoot using their membership number and current password. • U6:View Member Details: A Member views a subset of their details stored by iCoot, such as name, address and CreditCard details. • U7:Make Reservation: A Member reserves a CarModel when viewing its details. • U8:View Rentals: A Member views a summary of the Cars they’re currently renting. • U9:Change Password: A Member changes the password that they use to log on. • U10: View Reservations: A Member views a summary of Reservations that are not yet concluded, such as date, time and CarModel. • U11:Cancel Reservation: A Member cancels a Reservation that is not yet concluded. • U12:Log Off: A Member logs off from iCoot.

System use cases can be depicted on a use case diagram showing the actors and their associations with particular use cases – this helps us to see at a glance how the system will be used. A use case diagram for iCoot is shown in Figure 6.5.

147

148

Chapter 6

iCoot

U5: Log On

U1: Browse Index

U6: View Member Details

Customer U4: Search

U8: View Rentals

U2: View Results U3: View Car Model Details

U9: Change Password Member

U10: View Reservations

U12: Log Off

U11: Cancel Reservation

U7: Make Reservation

Assistant

Figure 6.5: A simple use case diagram for iCoot

On a use case diagram, each use case is shown as a number and a title, inside a bubble. The box around all the use cases represents the boundary of the system – we can put the system name just inside the box. Outside the system boundary, we show the actors, adding associations between the use cases and the actors that use them. A use case survey (Jacobson’s terminology for one of our non-UML artifacts) is an informal description of how a group of use cases ﬁt together: the kind of narrative that a developer might produce when walking sponsors through a use case diagram. A use case

Developer Perspective

survey allows sponsors to get a greater understanding of the use cases without any of the developers present.

Case Study iCoot use case survey Any Customer can look for CarModels in the catalog, either by browsing the CarModel index (U1) or by searching (U4). In the latter case, the Customer speciﬁes the Categories, Makes and engine sizes that they’re interested in. Either way, after each retrieval, the Customer is shown the results as a collection of matching CarModels (U2), along with basic information such as CarModel name. The Customer can then choose to view extra information about particular CarModels such as a description and an advert (U3). Customers who have become Members can log on (U5) and gain access to extra services. The extra services are: making a Reservation (U7), canceling a Reservation (U11), checking membership details (U6), viewing outstanding Reservations (U10), changing their log-on password (U9), viewing their outstanding Rentals (U8) and logging off (U12). Assistants are involved in the life cycle of Reservations, moving Cars to and from the reserved area, for example.

6.5.1 Specializing Actors Any actor can specialize (inherit behavior from) another actor. This adds more expressive power to the system use case model. For example, we may decide that Customer is an abstract concept that should be specialized by Member; once we have introduced this specialization, it makes sense to introduce the concept of NonMember too. It’s up to you whether to introduce inheritance between actors early, or late, or not at all. You (and your colleagues) need to be happy that each particular use of inheritance is beneﬁcial rather than confusing. Remember that all your sponsors must be able to understand the artifacts that you produce, from business modeling right the way through to static analysis, at least. This will help to ensure that you have understood the problem domain correctly and that you will deliver what the sponsors actually want. Will nonprogrammers be happy with NonMember is a kind of Customer? It’s up to you to decide.

149

150

Chapter 6

Case Study iCoot system actor list with inheritance Having decided to introduce inheritance for the iCoot actors, the ﬁnished system actor list looks like this (with one extra actor and the inheritance relationships in parentheses): • Customer: A person using a Web browser to access iCoot. • Member: A Customer who has presented their name, address and CreditCard details at one of our Stores; each Member is given an Internet password to accompany their membership number. (Specializes Customer.) • NonMember: A Customer who is not a Member. (Specializes Customer.) • Assistant: An employee at a Store who contacts Members to tell them about the progress of their Reservations.

We can modify our use case diagram to show inheritance relationships between actors in the same way as we would show them between classes, see Figure 6.6. The Customer class has been shown as abstract – its name is italicized – because no-one is literally a Customer; everyone is either a Member or a NonMember. (If you don’t have italics to hand, you can add the keyword {abstract} to the left or above the actor name.) Even though Customer is abstract, we can still associate it with a use case, to indicate that every kind of Customer is involved in that use case. As a side effect, some actors, such as NonMember, will have only indirect associations. Our diagram now shows specializations explicitly, but it’s useful to have the annotations in the actor list as well, in case the list is viewed separately.

6.5.2 Use Case Relationships As well as specialization between actors and associations between actors and use cases, there are three kinds of relationship between use cases themselves: specializes, includes and extends. These allow us to group related use cases; to decompose large use cases; to reuse behavior; and to specify optional behavior: • Specializes: Just like actors, use cases can inherit from each other. In order to avoid all sorts of complexity relating to the redeﬁnition of steps and the addition of extra ones, we can restrict ourselves to specializing abstract use cases. A (pure) abstract use case has no steps at all: its sole purpose is to group other use cases. For example, we might decide that U1:Browse Index and U4:Search are both varieties of the abstract use case U13:Look for CarModels.

Developer Perspective

NonMember

Customer

iCoot

U1: Browse Index U5: Log On

U4: Search

U6: View Member Details

U8: View Rentals

U3: View Car Model Details

U2: View Results

U9: Change Password Member

U10: View Reservations

U12: Log Off

U11: Cancel Reservation

U7: Make Reservation

Assistant

Figure 6.6: A use case diagram showing inheritance between actors

151

152

Chapter 6

• Includes: A use case that has some of its steps provided by another use case is said to include that use case. For example, U1:Browse Index includes, at some point in its behavior, all the steps of U2:View Results. Inclusion can be used to extract steps common to a number of use cases, or just to break a large use case down into more manageable chunks. • Extends: A use case that adds value to another use case is said to extend that use case. For example, when viewing results (U2), a Customer can choose to view details (U3). Extension allows us to add optional extras – often, these extras will appear at the end of the use case (hence the name), but they can also occur at the beginning, or somewhere in the middle. There is a fundamental difference between inclusion and extension: with inclusion, the source use case won’t work without the target, whereas, with extension, the source works perfectly well without the target. Going in the other direction, use cases that are included in other use cases may have independent existence – they may be executable directly via other routes. A use case that extends another use case, in contrast, will usually only exist as an extension. You’re unlikely to identify use case relationships on your ﬁrst pass through system requirements modeling. Beyond that, it is up to you to decide whether they’re really needed, and whether your sponsors will be able to appreciate them. If you do use relationships, as with other areas of object orientation, there will be many ways of decomposing use cases into inclusions, extensions and inheritance. No one way is correct – just try to develop a model that makes sense to you and your customers. Case Study iCoot use case relationships Our use case descriptions now look like those below. Abstract use cases and both ends of relationships have been identiﬁed – the rationale is that, seeing the annotation, the reader knows that they will have to look at other use cases to get a complete picture. • U1:Browse Index: A Customer browses the index of CarModels. (Specializes U13, includes U2.) • U2:View Results: A Customer is shown the subset of CarModels that were retrieved. (Included by U1 and U4, extended by U3.) • U3:View CarModel Details: A Customer is shown the details of a retrieved CarModel, such as description and advert. (Extends U2, extended by U7.) • U4:Search: A Customer searches for CarModels by specifying Categories, Makes and engine sizes. (Specializes U13, includes U2.) • U5:Log On: A Member logs on to iCoot using their membership number and current password. (Extended by U6, U8, U9, U10 and U12.)

Developer Perspective

Case Study (cont’d) • U6:View Member Details: A Member views their (censored) details stored by iCoot, such as name, address and CreditCard details. (Extends U5.) • U7:Make Reservation: A Member reserves a CarModel when viewing its details. (Extends U3.) • U8:View Rentals: A Member views a summary of the Cars they’re currently renting. (Extends U5.) • U9:Change Password: A Member changes the password that they use to log on. (Extends U5.) • U10: View Reservations: A Member views summaries of their unconcluded Reservations, such as date, time and CarModel. (Extends U5, extended by U11.) • U11:Cancel Reservation: A Member cancels an unconcluded Reservation. (Extends U10.) • U12:Log Off: A Member logs off from iCoot. (Extends U5.) • U13:Look for CarModels: A Customer retrieves a subset of CarModels from the Catalog. (Abstract, generalized by U1 and U4.)

Use case relationships can be shown on a use case diagram (see Figure 6.7). Inheritance between use cases is shown in the normal way, using a line with a white arrowhead. An inclusion is shown as a dashed, open-ended arrow from the including use case to the included one, labeled with the keyword <> (a word in guillemets in UML indicates a well-known concept). Therefore, U4:Search includes, at some point in its behavior, all the steps of U2:View Results. An extension is shown with a similar arrow from the extending use case to the extended one, labeled with <<extend>> – thus, U3:View CarModel Details is an optional extra for U2:View Results. Although we could reduce the two inclusions shown to a single inclusion, from U13 to U2, that would arguably be less clear. It would also mean that the abstract use case would have steps, something that we try to avoid. Finally, it would mean that U1 and U4 would not be able to control where the inclusion took place. In some cases, an extension is only allowed under certain conditions. We can show this by adding a UML comment (which looks like a piece of paper) detailing the conditions. A comment, which may contain any text, can be connected to the relevant point in the diagram with a dashed line, optionally terminated with a small circle that makes the join clearer. Conditions in UML are expressed as constraints (text in braces) and may be stated in natural language, in pseudocode or in UML’s formal Object Constraint Language (OCL). In Figure 6.7, natural language has been used, to reﬂect the fact that we’re at a relatively informal stage of development. As you can see, U7:Make Reservation is only allowed for Member objects who have logged on.

153

Chapter 6

iCoot >

<

U2: View Results <<extend>>

U10: View Reservations

U7: Make Reservation

te ex <<

nd

>>

U11: Cancel Reservation

tend> <<ex > d> ten ex < <

>

>

U5: Log On

Customer

d>> <<exten

Member}

U6: View Member Details <<extend>

<< ex ten d> >

a logged-On

NonMember U4: Search

U3: View Car Model Details {Customer is

U13: Look for Car Models

U1: Browse Index

<< inc lud e>>

<<extend>>

154

U12: Log Off

U9: Change Password

U8: View Rentals

Member

Assistant

Figure 6.7: The ﬁnal use case diagram for iCoot

Normally, a dashed line with an open-ended arrow in UML indicates a dependency – the source relies on the target in some way. The implication is that, if the target changes, the source is affected. For use cases, the use of dependency notation is historical and not strictly correct. For example, an extending use case doesn’t necessarily depend on the extended use case. Hence, with a use case diagram, we talk about use case relationships rather than dependencies. One last point about the notation: for inclusion, the subordinate use case is the target whereas, with extension, the subordinate is the source – this can cause confusion. Case Study iCoot use case survey (complete) Any Customer can look for CarModels in the catalog, by browsing the CarModel index (U1) or by searching (U4). In the latter case, the Customer speciﬁes the Categories, Makes and engine sizes that they’re interested in. Either way, after each retrieval, the Customer is shown the results as a collection of matching CarModels (U2), along with basic information such as CarModel name. The

Developer Perspective

Case Study (cont’d) Customer can then choose to view extra information about particular CarModels such as a description and an advert (U3). Customers come in two varieties, Members and NonMembers. Customers who have become Members can log on (U5) and gain access to extra services. The extra services are: making a Reservation (U7), canceling a Reservation (U11), checking membership details (U6), viewing outstanding Reservations (U10), changing their log-on password (U9), viewing their outstanding Rentals (U8) and logging off (U12). Assistants are involved in the life cycle of Reservations, moving Cars to and from the reserved area, for example. Browsing the index and searching for CarModels are two different ways of looking for CarModels (U13). In order to view CarModel details, a Customer must be viewing the results of looking for models (via the browsing or searching route). In order to reserve a CarModel, a Member must be viewing its details (NonMembers can’t make reservations, even when they’re viewing details). In order to cancel a Reservation, a Member must be viewing their outstanding Reservations.

6.5.3 System Use Case Details Once we’ve identiﬁed use cases and how they ﬁt together, we need to show the details. Since UML doesn’t specify what the use case details should include or how they should be arranged, a choice has been made here based on taste and experience. For Ripple, system use case details include: • • • • • •

The use case number and title. Whether the use case is abstract. relationships to other use cases. Any preconditions (conditions that must be satisﬁed before the use case is carried out). The steps themselves (where we can assume that the preconditions have been met). Any postconditions (conditions that are guaranteed after successful completion of the use case). • Any abnormal paths and what to do in each case (although the paths are abnormal, we include them if it’s important for us to specify the system’s reaction). • Any nonfunctional requirements that relate to this use case.

155

156

Chapter 6

Figure 6.8 shows the format of use case details used in this book; of all the items, only the number, title, preconditions, steps and postconditions are mandatory – the others can be left out if they’re empty. Number, Title (relationships) Preconditions Steps Postconditions Abnormal paths Nonfunctional requirements

Figure 6.8: Format for system use case details

Figure 6.9 shows four of the iCoot use cases in the speciﬁed format. These system use cases are signiﬁcantly more detailed than the business use cases we saw earlier. This reﬂects the fact that we’re now trying to be prescriptive rather than merely descriptive: we want to be precise about the services the system will provide in order to remove the guesswork from our analysts and designers. For the purposes of this book, we use a sequence of steps written in natural language. As a matter of personal preference, you may prefer to add a more algorithmic structure with conditionals and loops (if–then–else and repeat–until, for example). When writing use case details, it’s important that we specify the function of the system but not the way that function is delivered: for example, if we were to include steps such as 2. Customer clicks on the Details. . . button, we would be restricting the user interface designer. Unless it’s an absolute requirement, you should always try to use neutral words like select, initiate, indicate and display. Any reasonable designer will be able to make informed decisions about exactly how speciﬁc the requirements-gatherer is trying to be: for example, the designer may implement steps in a different order, or in parallel, as long as the end result is the same. We can see where something like this might happen in U5:Log On: the ﬁrst three steps could be implemented in any order and it would make no difference to anyone.

6.5.4 Preconditions, Postconditions and Inheritance Since we have considered inheritance between use cases, we must concern ourselves with how specialization affects preconditions and postconditions. (Although it’s recommended that you only inherit from abstract uses cases without steps, such use cases can still have preconditions and postconditions, as you can see from Figure 6.9.) Here are the rules:

Developer Perspective

U1: Browse Index. (Specializes U13, includes U2.) Preconditions: None. 1. Customer selects an index heading. 2. Customer elects to view CarModels for the selected index heading. 3. Include U2. Postconditions: None. U3: View CarModel Details. (Extends U2, extended by U7.) Preconditions: None. 1. Customer selects one of the matching CarModels. 2. Customer requests details of the selected CarModel. 3. iCoot displays details for the selected car model (makes, engine size, price, description, advert and poster). 4. If Customer is a logged-on Member, extend with U7. Postconditions: iCoot has displayed details of selected CarModels. Nonfunctional Requirements: r1. Adverts should be displayed using a streaming protocol rather than requiring a download. U5: Log On. (Extended by U6, U8, U9, U10 and U12.) Preconditions: Member has obtained a password from their local Store. 1. Member enters their membership number. 2. Member enters their password. 3. Since iCoot must enforce one logon for a Member, Member can choose to steal (invalidate and thus take over from) an existing session. 4. Member elects to log on. 5. Extend with U6, U8, U9, U10, U12. Postconditions: Member is logged on. Abnormal Paths: a1. If the membership number/password combination is incorrect, iCoot informs Member that one of the two is incorrect (for security, they’re not told which one). a2. If the membership number/password combination is correct, but Member is already logged on and they have not elected to steal, iCoot informs Member. U13: Look for Car Models (Abstract, specialized by U1 and U4.) Preconditions: None. Postconditions: Customer has been presented with summaries of retrieved CarModels.

Figure 6.9: Details for some iCoot system use cases

157

158

Chapter 6

1. When one use case specializes another, it inherits the parent’s preconditions as a starting point. Any new preconditions added by the child must only weaken the inherited ones (they’re combined using ‘or’). 2. For postconditions, the child’s starting point is the postconditions of the parent. Any postconditions added by the child must only strengthen the inherited ones (they’re combined using ‘and’). 3. Preconditions and postconditions added by children have no effect on the parent’s preconditions and postconditions. In the above list, rule 3 should be obvious from what you already know about objectoriented theory (children don’t affect the behavior of the parents), but rules 1 and 2 may be surprising to you. Informally, the reason we can only weaken preconditions and strengthen postconditions is that the child use cases have an obligation to readers of the parent use case not to spring any nasty surprises. For example, U13:Look for CarModels has no preconditions, so if U4:Search had a precondition that said Don’t do this on Tuesdays, anyone prevented from searching for car models on a Tuesday would have legitimate cause for complaint – ‘Excuse me, but according to ‘‘Look for Car Models’’, I can look at any time.’ A postcondition on a parent provides a guarantee to users and it wouldn’t be reasonable for the child to try and water down that guarantee. For example, U4:Search has a postcondition Customer has been presented with matching CarModels; it wouldn’t be reasonable for a child to add or a randomly-selected CarModel. Rules 1 and 2 imply that if a parent has Preconditions: None (there are no restrictions on when the use case may be applied), its children must also have Preconditions: None; if a parent has Postconditions: None (there are no guarantees about the outcome), the child can specify any postconditions it likes. In summary, when one use case specializes another, you must carefully consider the preconditions and postconditions of the parent.

6.5.5 Supplementary Requirements Most of the time, it’s possible to associate nonfunctional requirements with a particular use case. For example, a nonfunctional requirement of Adverts should be displayed using a streaming protocol rather than requiring a download ﬁts neatly with U3:View CarModel Details, the use case that makes adverts available to customers in the ﬁrst place. Nonfunctional requirements that don’t ﬁt with any use case can be recorded in a supplementary requirements document, as shown in Figure 6.10.

6.5.6 User Interface Sketches Thinking about the user interfaces for the system can help us to clarify the use cases. The interfaces can be brainstormed with our sponsors at an early stage and the results recorded

Developer Perspective

Supplementary Requirements -------------------------s1. The client applet must run in Java PlugIn 1.2 (and later versions). s2. iCoot must be able to cope with a catalog of 100,000 car models. s3. iCoot must be able to serve a million customers simultaneously with no significant degradation in performance.

Figure 6.10: Supplementary requirements for iCoot

as user interface sketches. These sketches should be regarded as a functional guide rather than a professional GUI design: they help us to identify and partition functionality in a way that can be implemented according to personal preference. For example, Sketch 1 in Figure 6.11 shows a user interface that allows the user to select one or more categories, manufacturers and engine sizes; on clicking the Retrieve button, we’re referred to Sketch 2 which shows a list of matching car models; clicking the < Go Back button returns us to Sketch 1; clicking the Details. . . button takes us on to Sketch 3 (not shown); and so on. (Obviously, these are not the original sketches that were hand-drawn on a whiteboard; they’re the versions that had been agreed by the end of the brainstorming session, which were then mocked up using a drawing package.) Since use cases and user interfaces both represent a partitioning of system functionality, it’s a good idea to maintain a clean mapping between the two, a mapping that survives through to implementation. For example, with iCoot, we have three broad categories of access: member access, nonmember access and assistant access. This suggests three separate user interfaces. Within each user interface, we should provide a window or a panel that corresponds to each of the use cases (the choice of window, panel, or some other widget is, of course, a design issue). For example, the sketches in Figure 6.11 show a notebook-style widget representing the nonmember interface. Nonmembers can search the catalog and browse the index, so each of these use cases is assigned its own page in the notebook. Both of these use cases include viewing the results – another use case that gets its own panel. We would expect the Search, Index and Results panels to be reused in the member interface. (For easier migration, the assistant interface will resemble the existing Auk interface.)

6.5.7 Prioritizing System Use Cases It’s a good idea, especially in the context of an incremental development process, to rank system requirements in order of their implementation priority. With use case modeling, the obvious thing to do is to rank the use cases – each can then be given a score to indicate its

159

160

Chapter 6

1: Search Index

category a category b category c category d engine size a engine size b engine size c engine size d

manufacturer a manufacturer b manufacturer c manufacturer d 2

Retrieve...

2: Search Index

car model a car model b car model c car model d car model e car model f car model g car model h car model i Details...

< Go Back 1

3

Figure 6.11: User interface sketches for iCoot

urgency. The priorities and urgencies can be used to help plan the rest of the development and any further increments. One useful scoring technique is trafﬁc lights: • Green use cases must be implemented in the current increment; failure to do so means that the project has failed to reach its minimum goals. • Amber use cases are optional for the current increment and should only be attempted once the green use cases have been completed (they’re added bonuses that we can use to impress our sponsors). Any amber use case that is incomplete by the delivery date must be dropped completely (partial implementations look unprofessional). • Red use cases won’t be implemented in the current increment, even if time permits: they’re outside the scope of the current increment and proper allowances are unlikely to have been made for them.

Developer Perspective

In practice, use case priorities (and urgencies) will be based not only on desirability, but also on how much of the system architecture and coding effort individual use cases will pull into the current increment: choosing priorities requires a certain amount of skill, experience and crystal-ball gazing. There’s nothing wrong with putting easier use cases ﬁrst: they will help us to learn more about the system as we iterate, with less risk. If you’re lucky enough to have time available at the end of the increment (after ﬁnishing the green use cases and all the amber use cases), you should: • Review the project status. • Finalize planning for the next increment (re-prioritizing unimplemented use cases, for example). • Do some unrelated work. • Have an ofﬁce party. • ...

Case Study iCoot use case priorities • Green: – U1:Browse Index – U4:Search – U2:View Results – U3:View CarModel Details – U5:Log On • Amber: – U12:Logoff – U6:View Member Details – U7:Make Reservation – U10:View Reservations • Red: – U11:Cancel Reservation – U8:View Rentals – U9:Change Password is both essential and simple (because it doesn’t involve rentals or reservations), so it goes at the top of the list; U5:Log On is essential before any of the member services can be made available, so it must appear before the member services; U6:View Member Details is attempted before U7:Make Reservation because it is simpler (reservations turn out to have a complicated life cycle); and so on.

U1:Browse Index

161

162

Chapter 6

Assigning priorities and urgencies to system use cases is another indication that we should develop for extensibility and reuse. Here’s a summary of how the trafﬁc lights ﬁt in with the other stages of development, after business modeling: • Green: System requirements, analysis, system design, subsystem design, speciﬁcation, implementation and testing should be complete for use cases in this group. • Amber: System requirements should be complete and analysis and system design should be complete, or nearly complete, for use cases in this group; subsystem design, speciﬁcation, implementation and testing are optional. • Red: System requirements should be complete for use cases in this group; analysis is optional; system design should support these use cases; subsystem design, speciﬁcation, implementation and testing should not be performed. Of course, ‘complete’ is a relative term in a spiral, iterative and incremental process.

6.6 SUMMARY In this chapter, we looked at: • The importance of specifying functional requirements (what the system must be able to do, such as ‘browse the catalog’) and nonfunctional requirements (how the system must run, such as the speciﬁc Web browsers that must be supported) in a requirements phase before any coding begins. • Modeling the business context and system functionality using high-level business use cases and identifying actors. • Modeling the system requirements with a complete use case model comprising use cases, use case diagram, supplementary requirements, user interface sketches, use case priorities and urgencies. Although communication diagrams and activity diagrams were considered optional at this stage, a glossary is always essential.

FURTHER READING Although it doesn’t cover UML, a mine of useful information on the theory of business process modeling is [Bustard et al. 00]. Ivar Jacobson’s original book on the Objectory method [Jacobson et al. 92], describes the original justiﬁcation for use cases and the kind of information that they should contain – it’s

Review Questions

always a good idea to see how an important technology came about in the ﬁrst place. Alistair Cockburn is a widely respected authority on use cases. As well as a book on the subject [Cockburn 00], Cockburn has his own web site at www.usecases.org. For more advice on communication diagrams and activity diagrams, Martin Fowler’s book [Fowler 03] is a good place to start. As ever, [OMG 03a] has comprehensive coverage of the notation.

REVIEW QUESTIONS 1. With reference to Figure 6.12, what are X1, X2 and X3? Choose only one option. (a) (b) (c) (d)

Roles. Prima donnas. Actors. Sticks.

UC1

X1

UC2 te ex

<<

UC3 >>

nd

X2

UC4

<< in clu de >>

X3

UC5

Figure 6.12: For use with Review Questions 1, 2 and 5

2. With reference to Figure 6.12, which of the following statements are true? Choose all options that apply. (a) (b) (c) (d)

X3 can interact with the system using UC4. X1 can interact with the system using UC1 and UC4. X3 and X1 are different kinds of X2. UC3 is an abstract use case with no steps of its own.

163

164

Chapter 6

PizzaBase Case Study The PizzaBase restaurant wants to automate the ordering of pizzas by customers. Each table will be ﬁtted with a touch-sensitive screen which customers can use to browse the pizzas on offer and select their choice. Two basic types of pizza will be offered: the Do-it-Yourself will have a base with tomato sauce only and then customers can choose any number of toppings, at a ﬁxed price per topping; the Prefab will come in several varieties, each with a ﬁxed set of toppings. Every pizza can be ordered with a deep crust or crispy base, and three sizes are available: 6 inch, 9 inch and 12 inch. Customers will also be able to order from a ﬁxed set of drinks, such as cola and lemonade ﬂavors, each in large or small size. Once customers have conﬁrmed their order, they will be shown the ﬁnal price and, thereafter, the screen will display the progress of their food as it is being prepared and cooked. At the end of a meal, payment will be made in the conventional way.

3. With reference to the PizzaBase case study, which of the following options are likely business use cases? Choose all options that apply. (a) (b) (c) (d) (e)

Customer pays for meal. Restaurant prepares meal. Customer sees progress of food. Customer chooses pizza. Customer selects drink from display.

4. Which of the following UML artifacts is used to show the steps involved in getting value from a system? Choose only one option. (a) (b) (c) (d) (e) (f)

User interface sketches. Glossaries. State machine diagrams. Use cases. Class diagrams. Deployment diagrams.

5. With reference to Figure 6.12, which of the following statements are true? Choose all options that apply. (a) (b) (c) (d) (e)

UC5 is a compulsory part of UC4. UC4 is an optional part of UC5. UC1 is unused. UC2 is an optional part of UC4. UC4 is a compulsory part of UC2.

Answers to Review Questions

B

A

C

Figure 6.13: For use with Review Question 6

6. With reference to Figure 6.13, what kind of objects are A, B and C? Choose only one option. (a) (b) (c) (d) (e) (f)

A is an entity, B is a controller, C is a boundary. A is a boundary, B is an entity, C is a controller. A is an entity, B is a boundary, C is a controller. A is a controller, B is an entity, C is a boundary. A is a boundary, B is a controller, C is an entity. A is a controller, B is a boundary, C is an entity.

ANSWERS TO REVIEW QUESTIONS 1. In Figure 6.12, X1, X2 and X3 are c. Actors. 2. With reference to Figure 6.12, all of the statements are true: (a) (b) (c) (d)

X3 can interact with the system using UC4. X1 can interact with the system using UC1 and UC4. X3 and X1 are different kinds of X2. UC3 is an abstract use case with no steps of its own.

3. The following options are likely business use cases: (a) Customer pays for meal. (b) Restaurant prepares meal. (d) Customer chooses pizza. 4. d. Use cases show the steps involved in getting value from a system. 5. With reference to Figure 6.12, the following statements are true: (a) UC5 is a compulsory part of UC4. (d) UC2 is an optional part of UC4. 6. With reference to Figure 6.13, d. A is a controller, B is an entity, C is a boundary.

165

7

Analyzing the Problem

I

n this chapter, we’ll be looking at the classical analysis phase and how to do it in a modern, object-oriented way. Analysis is an essential bridge between requirements capture and design, leading us from a clear statement of what the system must provide to a clear understanding of the objects that we’ll be dealing with. Once we understand the objects that we must deal with, we stand a much better chance of producing an elegant solution.

Learning Objectives Understand what analysis is.

•

Build a static analysis model.

Understand how dynamic analysis can help us to verify the static model.

Chapter Outline 7.1 Introduction 7.2 Why Do Analysis? 7.3 Overview of the Analysis Process 7.4 Static Analysis 7.5 Dynamic Analysis 7.6 Summary Further Reading Review Questions Answers to Activity 4 Answers to Activities Answers to Review Questions

7

168

Chapter 7

7.1 INTRODUCTION Analysis is about discovering what the system is going to handle, rather than deciding how to do the handling. We need to decompose a complex set of requirements into the essential elements and relationships on which we will base our solution. Analysis is our ﬁrst opportunity to get to grips with modeling the real world as objects. An analysis model has both static and dynamic parts. We can depict the static analysis model using a class diagram. A class diagram shows the objects that the system will handle and how those objects are related to each other. For the dynamic analysis model, we can use communication diagrams to demonstrate that our static model is feasible. As before, rather than all the intricacies of UML notation, you’ll see only the essential parts here: the parts that will sufﬁce for most purposes. There are two inputs to analysis: • The business requirements model (see Section 6.4) contains descriptions of the manual and automated workﬂows of our business context, described using business-oriented versions of actors, use cases, objects, the glossary and, optionally, communication diagrams and activity diagrams. • The system requirements model (see Section 6.5) contains an external view of the system, described as system-oriented versions of actors, use cases and use case diagrams, user interface sketches, an enhanced glossary and nonfunctional requirements. These inputs must be transformed into a model of the objects that will be processed by the proposed system, along with their attributes and relationships. These objects will exist within the system itself or at the system boundary, accessible via one or more interfaces. Most of the objects that we discover at this stage will correspond to physical objects or concepts in the real world (lower-level, solution-oriented objects won’t appear until design). Once we have a model of the system objects, we will put them through a veriﬁcation process to convince ourselves that they would support a solution.

7.2 WHY DO ANALYSIS? So, why do we do analysis in the ﬁrst place? Because analysis stops us designing a solution before we understand the problem. Although, in principle, we could jump straight into design and then implementation, hoping to draw out an understanding of the problem through trial and error, the analysis techniques we’ll look at here are much more efﬁcient. (Some well-chosen prototyping or proof-of-concept work is still permitted, if you feel the situation calls for it.)

Overview of the Analysis Process

We can’t expect to have a complete understanding of the problem from the business requirements model, because it describes existing practices: by adding software, we expect to be introducing new practices. Also, we have not separated manual workﬂows and automated (or potentially automated) workﬂows: for example, reserving a car model, as described in Chapter 6, involves both human-to-human and human-to-computer interaction. Even when we have a system use case model, our understanding of the problem is still incomplete, because the focus of use cases is external: use cases deal with the interactions between actors and the boundary of our system – the system itself is regarded as a black box, with only the outside visible. Use cases are imprecise: in order to make them easy to produce and easy to understand, use cases are written in natural language – therefore, they rely on our ability to understand language and to make certain assumptions. Business requirements modeling and system requirements modeling must still be done, of course: the former allows us to understand the business context; the latter forms a contract with our sponsors. Once we’ve completed static analysis, our sponsors will be able to conﬁrm that our understanding of the business objects is correct, before we let the objects inﬂuence our design. After dynamic analysis, we will be conﬁdent that our analysis objects support the required system functionality. In keeping with the philosophy of spiral development, dynamic analysis will also help us to build the static model. The static analysis model is also valuable when it comes to designing a database schema (for those business objects that need to be stored).

7.3 OVERVIEW OF THE ANALYSIS PROCESS In the case of Ripple, analysis has the following steps which you repeat until you and your sponsors are happy: 1. Use the system requirements model to ﬁnd candidate classes that describe the objects that might be relevant to the system and record them on a class diagram. 2. Find relationships (association, aggregation, composition and inheritance) between the classes. 3. Find attributes (simple, named properties of the objects) for the classes. 4. Walk through the system use cases, checking that they’re supported by the objects that we have, ﬁne-tuning the classes, attributes and relationships as we go – this use case realization will produce operations to complement the attributes. 5. Update the glossary and the nonfunctional requirements as necessary – the use cases themselves should not need updating, although perhaps they will need some correcting.

169

170

Chapter 7

The term ‘realization’ means ‘making real’. Operations discovered during use case realization should be disregarded during design – at this stage, we’re trying to build our conﬁdence not design the solution. You will need to show class diagrams, complete with attributes, to your sponsors, so that they can look for mistakes (those sponsors that understand the business probably understand it better than you do). A member of your team should summarize the information shown in the class diagram while the sponsors look on. This, coupled with the fact that class diagrams are relatively easy for nonprogrammers to understand, will elicit useful comments such as ‘Hang on, did you just say that nonmembers can’t reserve cars? You know, they can if they pay a deposit . . . ’ Deciding when to present the class diagram to sponsors is up to you (and your team), but you will generally do so at least twice: once to look for mistakes and once to verify that you’ve ﬁxed the mistakes. Generally, it’s not a good idea to show object operations or communication diagrams to your sponsors, because: • They add a lot of complexity. • They’re superﬂuous, as far as nonprogramming sponsors are concerned, because you’ve already demonstrated the dynamic behavior with the system use cases. • They imply code, something that’s deﬁnitely taboo for nonprogrammers. • They will be discarded before design anyway. Some sponsors, such as technical managers, may like to be shown a little of the dynamic analysis in order to increase their conﬁdence: ﬁne, but do this in separate, technical, meetings. In the rest of this chapter, we’ll look in detail at static and dynamic analysis, respectively.

7.4 STATIC ANALYSIS Static modeling involves deciding on the logical or physical parts of the system and how they should be connected together. Roughly speaking, it describes how we construct and initialize the system.

7.4.1 Finding Classes In the previous chapters of this book, classes were not identiﬁed systematically. Because we have now gone through the processes of business requirements modeling and system requirements modeling, we have a good source of candidate classes in the form of system use cases. Candidate classes are often indicated by nouns in the use cases. With a little practice, we can quickly cross out those nouns that represent:

Static Analysis

• The system itself, for example, ‘system’ or ‘iCoot’: As far as we’re concerned, the system is just a boundary for the development effort. • Actors, for example, Assistant or Head Ofﬁce: An exception to this is when we need to store information about an actor internally (for example, for Member, we need to store a password). Most of the time, actors are anonymous driving forces for our boundaries. • Boundaries, for example, ‘customer applet’ or ‘head ofﬁce link’: At this stage, we’re trying to identify business-related objects with interesting information and behavior. Boundaries are particular pieces of software that allow actors to get at our objects. • Trivial types (for example, strings and numbers): We can assume that these will be provided by the implementation language or its libraries. Short descriptions for the candidate classes that are left after this ﬁltering process should be added to the glossary. If you can’t write a short description for any class, maybe you’re expecting it to represent too much: consider splitting it into more than one class.

7.4.2 Identifying Class Relationships Once we have a list of candidate classes, we can try to draw relationships between them. There are four possible types of relationship: • Inheritance: A subclass inherits all of the attributes and behavior of its superclass(es). • Association: Objects of one class are associated with objects of another class. • Aggregation: Strong association – an instance of one class is made up of instances of another class. • Composition: Strong aggregation – the composed object can’t be shared by other objects and dies with its composer. Inheritance is a different kind of relationship to the other three: inheritance describes a compile-time relationship between classes while the others describe a run-time connection between objects. According to the UML standard, all run-time relationships come under the umbrella term association. However, most people use the term ‘association’ to mean ‘an association that isn’t aggregation or composition’. Choosing between relationships can be tricky – you need to use intuition, experience and guesswork. During analysis, you should expect the frequency of these kinds of relationship to be: association > aggregation > inheritance > composition As far as design and implementation are concerned, the differences between association, aggregation and composition can be difﬁcult to spot.

171

172

Chapter 7

7.4.3 Drawing Class and Object Diagrams A class diagram shows us what classes exist and how they’re related. (Ofﬁcially, class diagrams can also show attributes and operations, but that requires a lot more space.) In the case of aggregation, composition and association, the class diagram shows permitted run-time relationships rather than actual ones. Figure 7.1 shows a UML class diagram for iCoot. Every class is represented as a box with the class name inside (in bold, if not drawing by hand). If the class is abstract, the class name is italicized. If you’re labeling an abstract class by hand, you can add the keyword {abstract} above or to the left of the class name instead of using italics. 1..*

Car details

*

Rental

taken out by

Car Details

1 example of

Make

*

0..1

*

1

1..

rented under

*

Car Model Details

Customer

1 1

makes

*

details

Car Model

*

classified as 1

*

*

sold by

Reservation

*

1..

*

1..

guaranteed by

logs in with

lives at

1 1

Category

NonMember

1

Member

Vendor

Credit Card

1

Internet Account

1 Address

Figure 7.1: An analysis class diagram for iCoot

Relationships between classes are shown as lines with various annotations. Even without particular knowledge of UML, its easy to pick information out of a class diagram, just from the text. For example, we can see that ‘A Car can be rented under a Rental’, ‘A Rental can be taken out by a Customer’, and so on. Although the relationships on a class diagram are usually drawn between classes, the run-time relationship is actually between objects: for example, according to Figure 7.1, we would expect to see instances of Car connected to instances of Rental at run time. UML allows us to draw run-time objects as well as compile-time classes, as shown in Figure 7.2 Although UML allows us to mix classes and objects on the same diagram, people generally use the

Static Analysis

aRentalAgreement anAlphaRodeo:Car

aCar2 fredasAgreement:RentalAgreement

theYellowMolesChoice:Car

aCar1

Figure 7.2: Depicting objects in UML

term class diagram if there are no objects and object diagram if there are no classes (it’s up to you what you call a diagram that has both). In object diagrams, objects are shown as boxes connected by links – the links are ‘realized’ associations. So that we can easily spot the difference between classes and objects, object labels are underlined. As well as the object’s name, a label can include the object’s class, after a colon, as in fredasRental:Rental. We can show either the object’s name, or the object’s class, or both. If we show only the object’s class, we have to include the colon, to distinguish a class name from an object name, as in :Rental. Figure 7.2 shows two rentals: under the ﬁrst, aRentalAgreement, aCar2 and anAlphaRodeo have been rented; under the second, fredasRentalAgreement, theYellowMolesChoice and aCar1 have been rented. As you can see, object names correspond to the kind of variable names we would use in our programs. Object diagrams are useful for illustrating a particular run-time scenario, but they’re optional. For clarity, we would prefer to avoid putting classes and objects on the same diagram.

7.4.4 Drawing Relationships Figure 7.3 shows how inheritance is depicted on a class diagram: a white ﬁlled arrowhead on a solid line is drawn from the subclass to the superclass. In order to emphasize hierarchies of

173

174

Chapter 7

Car

Sports Car

Car Saloon Sports Car

Saloon

Figure 7.3: Depicting inheritance in UML

Car

Engine

Figure 7.4: Depicting aggregation in UML

Car

Body

Figure 7.5: Depicting composition in UML

Car

Driver

Figure 7.6: Depicting association in UML

subclasses, the arrows can be combined in the style shown on the left. Thus, SportsCar and Saloon are both subclasses of Car. Aggregation is drawn as a line between two classes with a white diamond on the aggregator’s end. So, Figure 7.4 shows that an Engine is part of a Car. Composition is drawn in a similar way to aggregation, but with a black diamond on the composer’s end – see Figure 7.5, which shows that a Body is always part of the same Car. Association is shown as an undecorated line (see Figure 7.6). Thus, a Driver is associated with a Car, but the Driver is not part of the Car (that would be aggregation) and the Driver is not always part of a single Car (that would be composition). When developing your analysis class model, for the sake of simplicity, make sure that information can only be deduced in one way. For example, the iCoot class diagram shown in Figure 7.1 allows us to calculate how many Cars a particular member is renting, from the

Static Analysis

class, the rented under association and the taken out by association. Thus, it would be redundant to show a hasOutForRent association from the Member class to the Car class, even if our use case model implied that such an association were needed.

Rental

Multiplicity All relationships except inheritance can indicate at either end the number of run-time objects that are allowed to take part in the relationship (the multiplicity of the relationship): • • • • •

n:

Exactly n. m..n: Any number in the range m to n (inclusive). p..*: Any number in the range p to inﬁnity. *: Shorthand for 0..*. 0..1: Optional.

For composition, the multiplicity at the composer’s end is always 1 because, according to the UML rules, a composed object can’t be shared among composites – thus a multiplicity would be redundant in this case. In other cases, if no multiplicity is shown, we must assume that it has not been speciﬁed, or that it is simply not known at this stage. It would be wrong to assume that a missing multiplicity implied some default value, such as 1. Looking at Figure 7.7, we can deduce the following: • • • • • • •

A Car has one Engine. An Engine is part of one Car. A Car has four or ﬁve Wheels. Each Wheel is part of one Car. A Car is always composed of one Body. A Body is always part of one Car and it dies with that Car. A Car can have any number of Drivers.

Body 1

Driver

*

1..

*

1 Car 1

Passenger 0..7

4..5

Engine Wheel

Figure 7.7: Depicting multiplicities in UML

175

176

Chapter 7

• A Driver can drive at least one Car. • A Car has up to seven Passengers at a time. • A Passenger is only in one Car at a time. Even though the notation we’ve seen so far appears to be unambiguous, things are never that easy, because we’re still dealing with natural language and assumptions about the real world that the reader will make. For example, the association between Car and Passenger, rather than being read as shown in the list above, could be read as the following pair of statements: • For each Car, there are only seven Passengers that could ever be in the Car. • A Passenger is only allowed in one particular Car. You may well argue that our initial interpretation is more obvious, based on what the modeler is likely to be trying to say (although we frequently talk about how many people a car can carry, when would we ever want to dictate who was allowed to ride in a particular car?). If the distinction is not obvious, or we wish to be pedantic for the sake of it, we can provide a verbose description – a class survey, if you like. Or, we could provide further annotation for our diagram using comments. The difference between aggregation and composition is subtle. In Figure 7.7, why is the Engine aggregated while the Body is composed? The differences relate to object sharing and object lifetimes. Recall that a composed object can never be part of more than one composite and dies with the composite, while an aggregated object can be shared and can outlive its aggregator. Although a car trundles out of the factory with a brand new engine inside, the engine may later be replaced, because it’s worn out, so the engine doesn’t necessarily die with the car; in contrast, the body of the car is an intrinsic part of the car – it’s the soul of the car, if you like, you can’t destroy the car without destroying the body (but you could always take the engine out ﬁrst). The issue of sharing is not important in this example: although the body could never be part of two cars (not legally, anyway), the engine couldn’t either. All this may seem a little confusing. The existence of composition in UML is really just a result of the properties of programming languages like C++. In C++, an object can be part of the same piece of memory as another object: in this case, the sub-object certainly dies with the larger object. This language property also leads to the second part of the rule: a composed object can’t be part of two objects at the same time. Even if the composed object is separate, we have no garbage collector in C++, so the composer may want to delete the composed object when the composer itself is deleted, which reinforces the danger of sharing. It would be neater if UML also required the composite to create the composed object (that would work well for our body example); however, that would prohibit us from creating the composed object ﬁrst and then passing it into the composite’s constructor.

Static Analysis

The fact that composition is really a programming issue implies that it is more common in design than analysis. As a matter of analysis style, it’s recommended that you use aggregation when there is an obvious part–whole relationship (as in the engine example) and save composition for those rare cases when there is an obvious shared lifetime (as in the body example). For design purposes, composition is also useful when you wish to add behavior to an object by hiding a delegate object inside, rather than by inheriting from another class (we saw an example of this in Chapter 3 with Stack and LinkedList). From an implementation point of view, one-to-many and many-to-many relationships often result in the use of collection objects (lists, trees, sets, etc.) at run time. For example, Car may employ some kind of List object to hold on to its passengers. One advantage of using multiplicities on class diagrams is that we do not need to be speciﬁc about such messy implementation details until much later.

Association Labels, Roles and Comments All relationships, except inheritance, can be given an association label, indicating the nature of the association. If it’s not obvious which way the association name should be read, a black arrowhead can be used. For example, in Figure 7.8, we can see that there is at least one Wheel that turns the Car.

Spare wheel is required by law 1

carrier

spare

Wheel

1

Car 1

1..2 turns 2 rear

Wheel

Wheel

Figure 7.8: Association labels, roles and comments in UML

As well as association names, we can show roles. A role indicates the part played by an object in the association – the role is shown as a label near the object that plays the role. For example, Figure 7.8 indicates the following roles: • A Car has one Wheel acting as a spare.

177

178

Chapter 7

• The spare Wheel has one Car acting as its carrier. • A Car has two rear wheels. In principle, association names and roles can be combined on the same association, but most of the time they should be considered alternatives (in order to avoid clutter). Figure 7.8 also shows a UML comment, an arbitrary piece of text enclosed in an icon that looks like a piece of paper, connected to the relevant part of the diagram by means of a dashed line. If the target of the dashed line is unclear, we can put a small white circle with a black border at the end – this is useful when the target is another dashed line, for example. A comment, which can appear on any diagram, can be used to provide extra information that would be difﬁcult or messy to show using other UML notation.

7.4.5 Attributes An attribute is a property of an object, such as its size, position, name, price, font, interest rate, or whatever. In UML, each attribute can be given a type, which is either a class or a primitive. If we choose to specify a type, it should be shown to the right of the attribute name, after a colon. (We might choose not to specify attribute types during analysis, either because the types are obvious or because we don’t want to commit ourselves yet.) Attributes can be shown on a class diagram by adding a compartment under the class name. To save space, we can document them separately instead as an attribute list, complete with descriptions. If we were using a software development tool, we would expect to be able to zoom in to see attributes (and their descriptions) or zoom out to see class names only. If you can’t provide a short description for an attribute at this stage, perhaps it should be several attributes, or even a class in its own right.

Engine capacity horsePower manufacturer:String numberOfCylinders fuelInjection:boolean

Figure 7.9: Depicting attributes in UML

Figure 7.9 shows the attributes of an Engine: capacity, horsePower, manufacturer, numberOfand fuelInjection. This diagram is speciﬁc about the type given to manufacturer (String) and fuelInjection (boolean); the implication is that we’ve decided that we will never Cylinders

Static Analysis

be interested in ﬁner details of the manufacturer – address, for example – or the particular variety of fuelInjection (we just want to know whether or not it’s there). As soon as we start showing attribute types, we open a can of worms: what is a String? What is a boolean? If the type is the name of one of our own classes, there’s no problem. Other than that, we don’t want to tie ourselves to a particular programming language or set of libraries. Therefore, it’s recommended that you stick to ubiquitous primitives (such as int, boolean and ﬂoat) and one or two obvious classes (for example, it’s fairly obvious that String indicates an object containing a sequence of characters). Although UML does allow us to deﬁne our own primitives in language-independent notation – Integer, Real and Boolean, for example – you might like to avoid using this facility because, when you come to design, you will have to be language-speciﬁc. (Another reason to avoid this issue is that, in Java, types such as Integer are classes, not primitives.) You may also like to avoid using the array type, which is usually a cross between an object and a primitive, even though most object-oriented languages support it. The reason is that your classes are likely to be more elegant if you use collection classes such as List and Set exclusively. During the design phase, you may ﬁnd yourself using arrays more, but you will need to be careful not to compromise good style for the sake of a slight improvement in performance. For simplicity, you should avoid including derived attributes in your artifacts. For example, a circle’s attributes include radius, diameter, circumference and area. However, we could store any one of these attributes and calculate the rest at run time, so we only need to show one of the four attributes on our class diagram. In this case, radius seems the obvious choice, because it’s probably going to be accessed more than the others (so we’d rather not calculate it) and the other attributes can be calculated using multiplication (which is faster than division). As far as UML is concerned, attributes and associations (all three varieties) are just properties of a class. In other words, every attribute can be shown as an attribute or as an association with the attribute’s name as the role (although an association to a primitive value or an array would look odd). This means that we can add multiplicities to attributes, after the type name, as in *, for a multi-valued attribute, or [0..1], for an optional attribute. This is UML’s way of avoiding the thorny issue of whether we should show an attribute or an association in any particular case. In this book, multiplicities won’t be shown for attributes, except in the case of optional attributes. Figure 7.10 shows a full set of attributes for the analysis objects discovered while examining the iCoot system use cases. For completeness, some of the attributes shown are from the full Coot system (totalAmount, for example). In order to avoid manipulating images and video in the system, adverts and posters are simply attributes that specify a location elsewhere (using a URL, for example).

179

180

Chapter 7

Car

CarDetails

name

Member inGoodStanding: boolean number:String

Address house street county postCode

Rental number startDate dueDate totalAmount

Category name

barCode:String numberPlate vin Make

CarModelDetails engineSize description advert poster

CarModel name price

travelled dateLost[0..1]

Vendor name

Internet Account password:String

CreditCard number:String type:String expiryDate

Customer name phoneNumber: String amountDue:int

Reservation number:String timestamp state

NonMember driversLicense: String

Figure 7.10: Attributes for iCoot

The dateLost attribute is optional (indicated by its [0..1] multiplicity): if the Car is lost, we record the date it was lost, otherwise we record nothing. In programming terms, we could use a null pointer to indicate that a particular Car is not lost. If an optional attribute has a primitive type, such as int, we have to reserve one value to indicate ‘no value here’: for example, our model might allow us to set aside 0 or -1. Occasionally, you’ll ﬁnd attribute multiplicities to be useful, but don’t overuse them. For example, only one attribute in this entire book (dateLost) required a multiplicity (even then, it could be argued that a more general notion of ‘car state’ would have proved better in the long run). As we saw earlier, UML allows us to draw run-time objects as well as compile-time classes. Figure 7.11 shows how we can specify run-time attribute values on object diagrams.

aRental startDate = 2004/06/23 dueDate = 2004/07/22 totalAmount = €1500

Figure 7.11: Depicting attribute values in UML

Static Analysis

Car

color 1

1

Color

Car color:Color

Car

Red Car

Green Car

Car

1

Color

Figure 7.12: Choosing between attributes and relationships

Attribute or Relationship? Often, we need to make choices between alternative ways of modeling information. For example, how would you choose to model the color of a Car, for the beneﬁt of customers? Figure 7.12 shows four alternatives: 1. 2. 3. 4.

Introduce an aggregation between Car and a class called color. Add an attribute to Car called color, with type color. Introduce a subclass of Car for each possible color. Introduce a composition between Car and color.

All of these options make sense, although some of them may seem counter-intuitive. But how would we choose? The central issue is this: which modeling option ﬁts the situation best? Or, to put it another way, which is the most natural? As far as option 1 goes, it seems a little heavy-handed to say that the color is part of the Car. Option 2 seems pretty good: as far as Car buyers are concerned, the color is just one of a Car’s attributes. Option 3 deﬁnitely seems to be over the top – do we really want a new type of Car for every available color, especially when there might be dozens? Option 4 seems a little more sensible than option 1 – a Car comes out of the factory sprayed a particular color; even if we change the color later, the original color is likely to remain underneath. Overall, option 2 would seem to be most appropriate in the context of buying cars.

181

182

Chapter 7

But would the choice be different if we were modeling cars on behalf of a car manufacturer? In that case, the manufacturer of the paint would probably become important – we would need to know where to get more if we ran out. So, we would need to model the color as a separate class with its own relationships and attributes; therefore, option 4 might be the best choice. Could we ever justify option 3? Perhaps. For example, if we were psychologists seeking to model the effect of a car’s color on the behavior of the driver – maybe red cars incite dangerous driving while green cars encourage caution. In that case, red and green cars may be sufﬁciently different to justify modeling them as separate classes. The moral of this story is that the analysts must choose whichever representation seems to ﬁt the current situation best: there is no correct answer. The best advice is probably not to worry about the philosophy too much. Instead, use common sense, experience, intuition, spirals and iterations to press forward to a successful implementation. In order to avoid confusion, you should ignore the fact that UML doesn’t distinguish between an attribute and an association-plus-role. Be guided by your model: if it seems to be an attribute, draw it as an attribute; if it appears to be an association, draw it as an association.

7.4.6 Association Classes Occasionally, an association has some information or behavior related to it. An association class can be introduced alongside the association, as depicted in Figure 7.13. This diagram indicates that a CarModel can be associated with any number of Customer objects and a Customer can be associated with any number of CarModel objects. For each link, there is a corresponding Reservation object that has a number, time-stamp and state. There is no name given to the association in this case, because it is implicit in the name of the association class. Reservation number:String timestamp state:ReservationState

CarModel

Customer

Figure 7.13: An association class from iCoot

An association class represents attributes and operations that exist only because the association exists: the attributes and operations are not tied to the objects at either end of the association. In the example above, when a customer makes a reservation, a new link is created at run time between the Customer and the corresponding CarModel. During requirements

Static Analysis

capture and analysis, it was determined that the reservation number, time-stamp and state must all be recorded. However, there is no sense in which these attributes ﬁt with the Customer or with the CarModel, they’re somewhere in between. Therefore, an association class is appropriate. When we come to design, we will have to replace association classes with something more concrete, because they’re not supported directly in most programming languages. However, they are useful during analysis.

7.4.7 Tangible versus Intangible Objects Often, you will ﬁnd yourself modeling an intangible object, such as a product described in a catalog, and a separate tangible object, such as the actual item that is delivered to your door. The object in the catalog describes the properties of something that you could order from the supplier, but which hasn’t necessarily been manufactured yet. The object that arrives at your door deﬁnitely has been manufactured, it is an instance of the type of product described in the catalog. Typically, there are many tangible objects for each intangible one. It is a common mistake to model tangible and intangible pairs as a single object. For example, if we were writing a sales system for a car dealership, we would discover during analysis that we were dealing with ‘catalogs’ describing the cars available for sale, ‘cars’ that we sell to customers, and ‘customers’ that buy the cars. It would be easy to conclude that we should produce the three concrete classes shown in Figure 7.14. However, in reality, we have two ‘car’ concepts here: the car that appears in the catalog is intangible, it describes the features of all cars of that type, but there may not be any such car in existence; in contrast, the car that is owned by a customer is tangible, it deﬁnitely exists, because it can be driven – it is separate from any similar car owned by a different customer.

Getting it Wrong To reinforce the issue of tangibility, let’s assume that, as well as selling cars, our dealership services customer cars on request. The information relating to sales includes: • modelNumber: This identiﬁes the production process for making a car of this type.

Customer

owns

Car Catalog

describes

Figure 7.14: Buying cars

183

184

Chapter 7

• availablecolors: The colors that cars of this type can be painted before they leave the factory. • numberOfCylinders: The number of combustion cylinders in the engine ﬁtted to this type of car. The information relating to services includes: • owner: The registered owner of the vehicle. • vehicleIdentiﬁcationNumber: A unique number stamped on a plate and riveted to the car’s body when it is manufactured that identiﬁes the car for registration purposes and helps the police to trace the owners of stolen cars. • mileageAtLastService: The number of miles the car had driven at the time of its last service – this allows us to calculate how many miles it’s covered since it was last serviced. Using the single concept of Car shown in Figure 7.14, we would have no choice but to put all of these attributes on one class, as shown in Figure 7.15. From what you already know about object modeling, the introduction of a class with two distinct sets of responsibilities should be ringing alarm bells already – such classes are said to have weak cohesion: their responsibilities do not form a single clump.

Car modelNumber availableColors numberOfCylinders owner vehicleldentificationNumber mileageAtLastService

Figure 7.15: A Car class, showing its attributes

Let’s suppose that we’re going to offer for sale the ‘Alpha Rodeo 156 2.0’: we would have to create a Car and set its attributes appropriately. This would give us the situation shown in Figure 7.16. (Possible attribute values have been shown as a list enclosed in braces – not strictly UML, but convenient for our purposes.) Now let’s suppose that a customer brings their ‘Alpha Rodeo 156 2.0’ in for its ﬁrst service. We now have two choices: we can create a new Car to represent the car owned by this particular customer, giving us the situation shown in Figure 7.17A, or we can use the existing Car, giving us the situation shown in Figure 7.17B.

Static Analysis

aCar:Car modelNumber = "Alpha Rodeo 156 2.0" availableColors = {red, green, silver} numberOfCylinders = 4 owner = vehicledentificationNumber = mileageAtLastService =

non-standard notation

Figure 7.16: A car for sale

A

aCarToo:Car

aCar:Car

modelNumber = "Alpha Rodeo 156 2.0" availableColors = {red, green, silver} numberOfCylinders = 4 owner = vehicledentificationNumber = mileageAtLastService =

modelNumber = "Alpha Rodeo 156 2.0" availableColors = {red, green, silver} numberOfCylinders = 4 owner = fredaBloggs vehicledentificationNumber = "VN19358" mileageAtLastService = 18036

B aCar:Car

modelNumber = "Alpha Rodeo 156 2.0" availableColors = {red, green, silver} numberOfCylinders = 4 owner = fredaBloggs vehicledentificationNumber = "VN19358" mileageAtLastService = 18036

Figure 7.17: Servicing a car

If we choose option A, half the attributes remain unused on the ﬁrst car object and we have redundant information on the second object. If we choose option B, we can only service one ‘Alpha Rodeo 156 2.0’ at a time (otherwise information about the ﬁrst one we started servicing will be lost). Although the original model shown in Figure 7.14 appeared natural and reasonable, the practical implications show that the model is nonsense. What we should have spotted early on is that we have an intangible concept, responsible for the ﬁrst set of attributes, and a tangible concept, responsible for the second set.

185

186

Chapter 7

Getting it Right Let’s discard the model shown in Figure 7.14, replacing it with a new intangible concept called CarModel and a tangible concept called Car. This gives us the class diagram shown in Figure 7.18. Now we can place the attributes modelNumber, availableColors and numberOfCylinders on CarModel, while owner, vehicleIdentiﬁcationNumber and mileageAtLastService remain on Car. Applying this new model to the example scenario we saw earlier, we create the run-time objects shown in Figure 7.19. Here we have one CarModel representing the ‘Alpha Rodeo 156 2.0’ and two Car objects representing the separate instances of this type of car that have been brought in for service. With this new model, we don’t care how many cars of Customer

owns

Car example of

Catalog

describes

CarModel

Figure 7.18: A tangible car and an intangible car model

alpha156:CarModel modelNumber = "Alpha Rodeo 156 2.0" availableColors = {red, green, silver} numberOfCylinders = 4

ex am pl e

of

fredasCar:Car example of

owner = fredaBloggs vehicledentificationNumber = "VN19358" mileageAtLastService = 18036

tomsCar:Car owner = tomThumb vehicledentificationNumber = "VN13788" mileageAtLastService = 99312

Figure 7.19: Car models for sale and cars that have been sold

Static Analysis

a particular type are sold, nor how many are returned for service, nor how many are serviced simultaneously, because the model can cope, logically and concisely, with all possibilities. Usually, a single intangible object will give rise to many tangible ones. Also, the attributes of intangible objects tend to be ﬁxed, while those in tangible objects tend to change over time. In our example, there is one CarModel, representing any number of ‘Alpha Rodeo 156 2.0’ Car objects that we might sell. The attributes of CarModel do not change over time (the manufacturer may occasionally change the speciﬁcation, by adding new colors, for example, but that doesn’t happen very often). The model shows two Car objects, each representing a particular ‘Alpha Rodeo 156 2.0’ that’s been returned for service, at least once, by its owner. The attributes of Car are changeable: the owner changes when the car is sold and the mileageAtLastService changes every time the car is serviced. The vehicleIdentiﬁcationNumber doesn’t change, but this is a special case of an identity attribute, a property that distinguishes this object from all other similar objects, throughout its lifetime.

ACTIVITY 4

Before we leave the subject of tangibility, consider a video rental system. Which of the two class diagrams shown in Figure 7.20 is correct, do you think?

Customer

rents

Video

Member

Customer

reserves

rents

Video

recording of

Member

reserves

Movie

Figure 7.20: Tangible and intangible objects from a video rental system

187

188

Chapter 7

7.4.8 Good Objects It’s all very well being able to draw class diagrams in perfect UML notation. It’s quite another thing to be able to ﬁnd good objects, attributes and relationships. In this book, the recommended starting point is nouns in use cases, a combination of the obvious with the valuable work of Jacobson. Beyond that, a lot of your success will come from hard thinking and growing experience. Domain experts (customers and colleagues) can help here, since they should be invited to comment on your class model. Using spirals and iterations will also help, a lot. If you ﬁnd yourself struggling to decide what should be an object and what shouldn’t, remember that objects are cheap – if something you’re thinking about or talking about sounds like an object, it probably is an object: don’t be afraid to put it down on paper, look at it and see what it can do – at this point, you’re a long way off the heavy investment of writing code. If you ﬁnd yourself thinking about the properties of an object, you have attributes; if you’re thinking about what an object should be able to do, you have operations. In this book, it’s recommended that you don’t worry too much about operations until dynamic analysis, but there’s no harm in noting some down in the meantime – dynamic analysis is used to verify that you have at least the operations you need in order to satisfy the use cases, but a few more won’t hurt. Another good trick, if you can’t seem to ﬁnd the right objects from the use cases, is to talk to an independent colleague about the business or the system in question. Ask them to note down everything you mention that seems to be an important concept, in exactly the same way as if they were taking notes at a lecture. This way, any unhelpful anxieties and prejudices that you have will be side-stepped. The practical advice you’ve seen on choosing between attributes and relationships, considering the tangible and the intangible, and identifying association classes should help you to avoid common pitfalls.

7.5 DYNAMIC ANALYSIS We perform dynamic analysis for the following reasons: • To conﬁrm that our class diagram is complete and accurate, so that we can ﬁx it sooner rather than later: this may involve adding, deleting or modifying classes, relationships, attributes and operations. • To gain conﬁdence that our modeling up to this point can be implemented in software: we’re not the only ones that should be conﬁdent before we proceed, our sponsors are just as important.

Dynamic Analysis

• To verify the functionality of the user interfaces that will appear in the ﬁnal system: it’s a good idea to partition access to the system into separate interfaces, along use case lines, before we dive into detailed design. According to Jacobson, the most important part of dynamic analysis is use case realization, i.e. making our use cases real by demonstrating how they can be implemented as collaborating objects. Use case realization has the following steps: 1. Walk through the system use cases, simulating the messages sent between objects and recording the results on communication diagrams. 2. Introduce operations on the objects that receive the messages. 3. Add classes to represent boundaries (system interfaces) and controllers (placeholders for complex business processes or for the creation and retrieval of objects), as necessary.

7.5.1 Drawing Use Case Realizations As we simulate the messages sent between our analysis objects, we need to record our results. UML communication diagrams and sequence diagrams are designed for this purpose. Although we can pretty much record the same information on communication diagrams and sequence diagrams, communication diagrams are better for use case realization because they’re simpler to produce and they focus on the objects and their connections, rather than on the order in which the messages are sent. We saw a simple business-level communication diagram in Section 6.4.4. Unlike our business requirements example, which had a rather free-form mix of actors, objects and systems, the analysis example (Figure 7.21) is cleaner. An informal description of the information shown in Figure 7.21, from a member of the iCoot team, might read as follows: A Member actor asks the MemberUI to reserve a particular CarModel; the Member is warned that there is a ﬁne if the corresponding car is not collected when it arrives; once the Member conﬁrms that they do wish to make a reservation, the MemberUI asks the ReservationHome to create a new Reservation, passing in the CarModel and the Member (which the user interface already has, as a result of logging on); ﬁnally, the MemberUI gets the number from the new Reservation and passes it to the Member. In general, analysis-level communication diagrams can show: • Actors interacting with boundaries (for example, the Member interacts with a MemberUI). • Boundaries interacting with objects inside the system (for example, the MemberUI interacts with a ReservationHome, a Member, a CarModel and a Reservation).

189

190

Chapter 7

m:Member c:CarModel

1:reserve(c) 1.1:fineWarning() 1.1.1:confirmReserve() 1.1.1.3:reservationNumber(n)

Member

r()

be

m

Nu

t ge

= :n

.2

1 1.

1.

:MemberUI

1.

1.

1.

1:

r=

cr

ea

te

(c ,m

)

r:Reservation

:Reservation Home

Figure 7.21: A communication diagram for U7:Make Reservation

• Objects inside the system interacting with boundaries to external systems (we might imagine an internal ReportGenerator object interacting with a HeadOfﬁce boundary, for example).

We do not need to show any business objects that lie outside the system, nor do we need to show actors that don’t interact with the system directly. Rather than two-way interactions, we can now use a more computer-oriented, client– supplier style: actors initiate interactions with boundary objects; boundary objects initiate interactions with system objects; system objects initiate interactions with other system objects, and boundaries to other systems.

Dynamic Analysis

7.5.2 Boundaries, Controllers and Entities A plain communication diagram shows objects as labeled boxes. For extra expressiveness, UML allows the developer to use icons, instead of boxes, to indicate the nature of the object. Figure 7.22 shows the UML meanings of the icons in Figure 7.21:

Actor

Entity

Boundary

Controller

Figure 7.22: Jacobson’s icons for communication diagrams [Jacobson et al. 92]

• Actor: A person (usually) or system (occasionally) existing outside the system. • Boundary: An object at the edge of the system, between the system and the actors. For system actors, boundaries provide a communication path. For human actors, a boundary means a user interface, capturing commands and queries and displaying feedback and results. Each boundary object usually corresponds to a use case, or a group of related use cases. More speciﬁcally, such a boundary usually maps to a user interface sketch (in which case it may be an entire interface or just a sub-window). It is quite reasonable for boundary objects to survive through to design. • Entity: An object inside the system, representing a business concept such as a customer, a car or a car model and containing useful information. Typically, entities are manipulated by boundary and controller objects, rather than having much behavior of their own. Entity classes are the ones that appear on our analysis class diagram. Most entities survive through to design. • Controller: An object inside the system that encapsulates a complex or untidy process. A controller is a service object that provides the following kinds of service: control of all or part of a system process; creation of new entities; retrieval of existing entities. Without controllers, our entities would become polluted with messy details. Since controllers are just a convenience for the beneﬁt of analysis, we do not expect many of them to survive through to design; an important exception to this is the idea of a home. A home is a controller that is used for the creation of new entities and the retrieval of existing ones. A home may also have utility messages as in carModelHome.ﬁndEngineSizes(). Since a home is such a clean concept, they often survive through to design.

191

192

Chapter 7

The RUP approach is to retain all controllers for the design phase, complete with all the operations that we discover during dynamic analysis. With RUP, there is no distinction between the analysis model and the design model – we simply start with the analysis model and enrich it again and again, until it is transformed into an implementable design model. For the purposes of this book, and in general, this approach is considered too optimistic. The valuable outputs from analysis are: • • • •

good entity objects with validated attributes; high-level boundary objects that mirror the use cases; conﬁdence that our model is correct; homes (ignoring any utility messages).

The designer should not have the opportunity to modify these fundamental outputs for the sake of implementation. The ﬂip side to this is that the analyst should not be considering programming details, such as how to implement object attributes, relationships or operations. It’s recommended that designers begin with a fresh class diagram that is seeded with the entity objects that were discovered during analysis. Selected boundaries and homes can then be added to this diagram, as appropriate.

7.5.3 Communication Diagram Elements Figure 7.23 shows the analysis communication diagram again, with annotations that explain the individual pieces of notation. The details of the notation are as follows: • • • • •

• • • •

Actors are shown in the same way as they are on use case diagrams. Objects are shown as labeled icons or labeled boxes. A line between two objects indicates a link, as on an object diagram. A message is shown as a sequence number (indicating the message’s position in the communication), a message name (in the usual format) and a parameter list (in parentheses). An open-ended arrow shows the direction the message is being sent (this arrow doesn’t have to be positioned at the end of the message, some developers place it below, for example). Labels, used to identify objects and parameters, can be shown as name, name:Type, :Type or literal (for example, carModel, m:CarModel, :CarModel, 10 or "abc"). Assignment of a reply value to a name can be shown like this: n = getNumber(). A conditional message can be shown as a guard (a condition in brackets), next to the message, for example, 4:[Only on Saturday]readPaper(). Iteration can be indicated with a * next to the part of the sequence number that iterates.

Dynamic Analysis

icon (or labeled box) m:Member

c:CarModel

link

actor

1:reserve(c) 1.1:fineWarning() 1.1.1:confirmReserve()

Member

1.1.1.3:reservationNumber(n) message r() be m u tN ge n= 2: 1. . 1 1. direction

:MemberUI

sequence number

1.

1.

1.

assignment

1:

r=

cr

ea

te

(c ,m

)

parameters (name, :Class, name:Class, or value)

r:Reservation name (optional)

class (optional)

:Reservation Home

Figure 7.23: Elements of a UML communication diagram

Although, in reality, no actor would literally ‘send a message’ to a boundary object, the message metaphor is a convenient and concise way of representing an interaction. Unlike object diagrams, object labels on communication diagrams are not underlined (partly because classes can’t be present to cause confusion). A few words about sequence numbers. Initially, a message is sent to an object, causing the execution of a method. Within that method, further messages can be sent: these are given sequence numbers that show the level of the interaction. For example, message 1 causes method 1 to execute; its ﬁrst message is 1.1, the second is 1.2, and so on. The ﬁrst message sent within method 1.2 is 1.2.1. As in Figure 7.23, when a communication diagram shows the implementation of a single top-level message (reserve), the ﬁrst number in the sequence is never higher than 1. Although this numbering scheme can lead to many levels of nesting (especially where we have a two-way interaction), it does at least allow us to be precise.

193

194

Chapter 7

If we’re dealing with concurrent communications on a single diagram, we can give each communication – or thread – a name, as part of the sequence number. For example, we might use a and b as the names of two separate threads; 2.2a and 2.2b would happen at the same time, while 2.3a would happen later on thread a.

7.5.4 Adding Operations to Classes Every message on a communication diagram corresponds to an operation on a class, so we should record the operations in order to have a complete set of use case realizations. Operations can be shown on a class diagram in a separate compartment below the attribute compartment, as shown in Figure 7.24. Alternatively, they can be documented as a separate operation list, to save space. The style you choose is likely to be dictated by whether you’re using a development tool, paper or a whiteboard; however you do it, be sure to include descriptions of the operations too. MemberUI reserve(:CarModel) confirm()

Reservation number:String timestamp state: ReservationState getNumber():String

ReservationHome create( :CarModel, :Member ):Reservation

Figure 7.24: Some operations from iCoot

The general form of a UML operation is: opName(paramName1:paramType1,paramName2:paramType2):ReturnType

Each parameter name, parameter type and return type is optional. (If the parameter name is omitted, the colon must be retained in front of the parameter type, to avoid confusion.) For an empty parameter list, the parentheses can be omitted, but they’re retained in this book, so that operations are instantly distinguishable from attributes.

7.5.5 Responsibilities Whenever you add detail to your objects, you should consider responsibility as described in [Wirfs-Brock and McKean 02]. This will help you to ﬁnd and correctly assign operations and

Dynamic Analysis

attributes. (Although you will have discovered many of the attributes during static analysis, during use case realization you will discover the need for more.) Whenever you discover some information or behavior that needs to be in the system, think to yourself ‘Which object is responsible for this?’ Information ends up as attributes (and objects); behavior ends up as operations (and collaborations). Taking the idea of responsibilities a step further, make sure that no object is responsible for more than one job (or role, if you prefer): if you have an object that is responsible for getting order details from a customer and processing the order, you probably want two objects (a boundary and a controller, perhaps). Objects with a single set of responsibilities are said to have strong cohesion, a desirable goal. Also, remember to think of your objects as clients (who ask questions and give commands) or suppliers (who provide answers and perform services). The alternative, two-way collaboration, leads to greater complexity and more difﬁcult maintenance. A supplier is said to be loosely coupled to its client. Two-way collaborations lead to strong coupling in both directions. The fact that each use case realization described in this book involves an actor as initiator talking to a boundary that talks to entities, should help you to produce client and supplier objects. Of course, going deeper into the system, a supplier may also be a client and a client may also be a supplier.

7.5.6 State Modeling Sometimes, an entity will have a life cycle that’s complex enough to be shown on a state machine diagram. For example, Figure 7.25 shows a model of the complex life of a Reservation in the iCoot system. In this diagram, a box with rounded corners indicates a state, with a label giving its name. An arrow indicates a transition to another state – the label on the arrow indicates the trigger that causes the transition. A black circle with an arrow coming out of it points to an initial state – a state into which an object can be born. An arrow pointing to a ringed black circle indicates that the source is a ﬁnal state – a state where an object can end its life. As a result of the start and stop states, Figure 7.25 shows that, when a Reservation is created, it is Waiting and, once it has reached the Concluded state, it plays no further part in the system. As for the other parts of the diagram, rather a lot is being said. Below is a full description – a state machine survey, perhaps. This description alone should convince you of the need for a diagram. When a Member reserves a CarModel over the Internet, the Reservation is initially Waiting to be processed by an Assistant (this is so the Customer can make a Reservation without the intervention of an Assistant). The Reservation becomes Notiﬁable if, some time later, an Assistant ﬁnds a suitable unreserved Car in the display area of the car park, or if one is returned by a Customer. In this case, the Car is moved to the reserved area.

195

196

Chapter 7

no car found no car arrives

Waiting

member cancels

car assistant returned finds car

member renews

Notifiable

member unreachable member cancels

assistant notifies member member cancels

Needing Renewal

member no-show

Collectable member collects

member cancels

member cancels

Storable

member unreachable

assistant stores

Concluded

Figure 7.25: State machine diagram for a reservation

If no Car becomes available for a particular Reservation within a week, the Reservation becomes NeedingRenewal: the Member must be contacted, by phone or in person, so that they can cancel the Reservation, or ask for it to be renewed for another week. If the Member cancels or can’t be contacted within ﬁve days, the Reservation is Concluded. Once a Reservation is Notiﬁable, the Member must be notiﬁed by an Assistant, in person or by phone, within three days; if the Customer can be reached, the Reservation is Collectable otherwise it becomes Displayable (a Car that was moved to the reserved area must be returned to the display area). Once a Reservation is Collectable, the Member must collect the Car within three days: if they do collect, the Reservation is Concluded; otherwise, the Reservation becomes Displayable. Once a Displayable Reservation’s Car has been put back in the display area, the Reservation is Concluded. At any time, the Member may cancel the Reservation over the Internet, by phone or in person. The system will keep Assistants informed as to the state of current (not yet concluded) reservations, so that they can take appropriate action. As well as behavior state machines, such as the one we’ve just seen, UML has protocol state machines. The latter are used to show the order in which messages can legally be sent to an object, but otherwise they’re similar.

Review Questions

7.6 SUMMARY In this chapter, we’ve looked at: • How to perform the analysis phase of software development. • How to build a static analysis model showing the business-oriented objects of our proposed system, along with their attributes and relationships, on a class diagram. • How dynamic analysis can improve and verify the static model, using communication diagrams, and how we can model complex life cycles, using state machine diagrams.

FURTHER READING Finding good objects, attributes and relationships comes partly from good thinking, partly from talent and partly from experience. Despite this being a rather philosophical activity, there are places to look for help. Three popular books that are worth investigating are [Fowler 96], [Larman 01] and [Martin and Odell 98]. Fowler gives many examples of analysis models from the real world. The other two books, although they cover a much broader range of topics, do address the issue of ﬁnding conceptual objects. For ﬁnding responsibilities, [Wirfs-Brock and McKean 02] is the foundation. The original ideas on breaking a system up into boundaries, controllers and entities are due to Jacobson [Jacobson et al. 92]. In this book, state machine diagrams are covered at a conceptual level only – Martin Fowler provides more detail in [Fowler 03] and the full picture can be found in the UML Speciﬁcation [OMG 03a].

REVIEW QUESTIONS 1. With reference to Figure 7.26, what do you think is the most likely implementation of the relationship between Car and Engine? Choose only one option. (a) (b) (c) (d)

A ﬁeld, of type Car, in Engine. A class called CarEngine with one ﬁeld of type Car and another ﬁeld of type Engine. A ﬁeld, of type Engine, in Car. A ﬁeld, of type Engine, in Car and a ﬁeld, of type Car, in Engine.

2. With reference to Figure 7.26, which of the following statements are true? Choose all options that apply. (a) A car always has the same body.

197

198

Chapter 7

Body 1

Driver

*

1..*

Car

1

1

Engine

1 Passenger

0..7

4..5

Wheel

Figure 7.26: For use with Review Questions 1 and 2

(b) (c) (d) (e) (f)

Some cars have spare wheels. A car has one engine, and engines are not shared between cars. All cars have either four or ﬁve wheels. A car must have at least one driver. Passengers cannot be drivers.

PizzaBase Case Study The PizzaBase restaurant wants to automate the ordering of pizzas by customers. Each table will be ﬁtted with a touch-sensitive screen which customers can use to browse the pizzas on offer and select their choice. Two basic types of pizza will be offered: the Do-it-Yourself will have a base with tomato sauce only and then customers can choose any number of toppings, at a ﬁxed price per topping; the Prefab will come in several varieties, each with a ﬁxed set of toppings. Every pizza can be ordered with a deep crust or crispy base, and three sizes are available: 6 inch, 9 inch and 12 inch. Customers will also be able to order from a ﬁxed set of drinks, such as cola and lemonade ﬂavors, each in large or small size. Once customers have conﬁrmed their order, they will be shown the ﬁnal price and, thereafter, the screen will display the progress of their food as it is being prepared and cooked. At the end of a meal, payment will be made in the conventional way.

3. With reference to the PizzaBase case study, which of the following is most likely list of attributes at the analysis stage? Choose only one option. (a) cola, base, price, size, lemonade, payment. (b) ﬂavor, variety, payment, ﬁnal, display, meal, tomato.

Review Questions

(c) progress, variety, ﬂavor, price, touchSensitive, size, drink. (d) base, price, variety, size, progress, ﬂavor. 4. With reference to Figure 7.27, which diagram is the best model of Pizzas in the PizzaBase restaurant? Choose only one option. specifies

Order

Pizza Menu

Diagram 1

lists Customer

orders

Pizza

Diagram 3 Order

specifies

Choice

kindOf lists

Menu

Diagram 2

Pizza

Figure 7.27: For use with Review Question 4

(a) Diagram 1. (b) Diagram 2. (c) Diagram 3. 5. With reference to the PizzaBase case study, which of the following is the most likely list of analysis classes? Choose only one option. (a) Payment, Order, Drink, Topping, Pizza, Order, Restaurant, Base, Sauce. (b) Customer, Table, Pizza, Topping, Drink, Restaurant, Order. (c) PizzaBase, Cola, Restaurant, Lemonade, Customer, Do-it-Yourself, Prefab, Table, Order. (d) Restaurant, Pizza, Topping, Display, Order, Payment, Touch. (e) Screen, Order, Offer, Topping, Size, Meal, Pizza, Restaurant. (f) Pizza, Customer, Cook, Table, Crust, Topping, Drink, Restaurant. 6. In UML, which diagrams are used to show messages sent between objects? Choose all options that apply. (a) Activity diagrams. (b) Object diagrams.

199

200

Chapter 7

(c) (d) (e) (f)

Communication diagrams. State machine diagrams. Sequence diagrams. Deployment diagrams.

7. With reference to Figure 7.28, which kind of icon would you use to represent a business object containing useful information? Choose only one option. (a) A. (b) B. (c) C.

B

A

C Figure 7.28: For use with Review Question 7

8. What is an association class? Choose only one option. (a) It describes the various kinds of relationship that can exist between classes. (b) It adds attributes and/or behavior to an association between two other classes. (c) It associates an object with the class of which it is an instance. 9. With reference to Figure 7.28, which kind of icon would you use to represent a communication path between systems or between a human and a system? Choose only one option. (a) A. (b) B. (c) C.

Answers to Review Questions

10. With reference to Figure 7.28, which kind of icon would you use to represent an object that coordinates a system process, creates objects or retrieves objects? Choose only one option. (a) A. (b) B. (c) C.

ANSWERS TO ACTIVITY 4 The lower option in Figure 7.20 is correct. If there’s any doubt in your mind, consider the placement of the attributes barCode and title. A title belongs to a Movie and every recording of it will have the same title. However, every physical Video copy of a Movie will have its own barCode. This illustrates that, sometimes, the name of an intangible class can be completely different to that of its tangible partner.

ANSWERS TO REVIEW QUESTIONS 1. The most likely implementation of the relationship between Car and Engine is c. 2. With reference to Figure 7.26, the following statements are true: a. A car always has the same body. c. A car has one engine, and engines are not shared between cars. d. All cars have either four or ﬁve wheels. 3. The most likely list of attributes at the analysis stage is d. base, price, variety, size, progress, ﬂavor. 4. With reference to Figure 7.27, b. Diagram 2 is the best model of Pizzas in the PizzaBase restaurant. 5. The most likely list of analysis classes is b. Customer, Table, Pizza, Topping, Drink, Restaurant, Order. 6. In UML, c. Communication diagrams and e. Sequence diagrams are used to show messages sent between objects. 7. I would use icon B to represent a business object containing useful information. 8. An association class b. Adds attributes and/or behavior to an association between two other classes.

201

202

Chapter 7

9. I would use icon C to represent a communication path between systems or between a human and a system. 10. I would use icon A to represent an object that coordinates a system process, creates objects or retrieves objects.

III

Designing the Solution

8

Designing the System Architecture

T

his chapter considers how to move from analysis into design and concentrates on the system architecture elements of design.

Learning Objectives Understand the steps involved in system design and how a system can be decomposed into physical and logical components.

•

Illustrate architectural decisions on UML deployment diagrams.

Understand the concurrency and security issues that arise in networked systems.

•

Understand how to partition a system and add partitioning decisions to a UML deployment diagram.

Chapter Outline 8.1 Introduction 8.2 Design Priorities 8.3 Steps in System Design 8.4 Choosing a Networked System Topology 8.5 Designing for Concurrency 8.6 Designing for Security 8.7 Partitioning Software 8.8 Summary Further Reading Review Questions Answers to Review Questions

8

206

Chapter 8

8.1 INTRODUCTION Analysis and design are very different ideas, although the boundaries are sometimes blurred. This blurring can happen intentionally, as in the case of RUP, or accidentally, resulting from poor software development. A clear separation between analysis and design is a good idea, to make sure that the problem is well understood before a solution is considered. With that separation in mind, analysis is about investigating the problem while design is about inventing a solution. Or, to put it succinctly, ‘Analysis = What; Design = How’. There are no strict rules for transforming an analysis model into a design model. Anyone who tells you otherwise is ignoring the human factors and creativity involved in software development – if we could be systematic, software development would be just like engineering, an idea that’s already been debunked in Chapter 5. The design process is driven by the need to produce a ﬁnished system; the experience of the team; reuse opportunities; personal preference. Once the designer has studied the requirements and analysis artifacts, he or she can start with a clean sheet of paper: we don’t care whether there is a close correspondence between analysis objects and design objects, as long as the design leads to an effective solution. During the design phase, we make certain technology choices (for example, programming languages, protocols and database management systems). We must decide how much impact we want these choices to have on our design. Our technology choices will inﬂuence the libraries, patterns and frameworks that are available to us and even the detailed UML notation that we use. The more general we can make our design, the less we are tied to a particular technology – this will reduce the need for developers to be experts in multiple technologies and it will protect us from technologies that become obsolete or unsupported. The downside of being general is that we may not get the maximum beneﬁt from any particular technology (reuse and performance being the important factors here). History suggests that individual technologies appear and disappear more frequently than the theories that underpin those technologies. For example, programming languages have included COBOL, Fortran, Pascal, Ada, Modula, PL/1, C, C++, Smalltalk, Eiffel, C# and Java, but the use of these technologies has been governed by only two theories: structured programming and object-oriented programming. Therefore, it’s reasonable to conclude that being general is safer than being speciﬁc. With respect to the reuse penalty of a general design, we must accept that some investigation will be needed to discover new reuse opportunities for each implementation. As for the performance penalty, we would hope that keeping our design general would not lose us more than, say, 10% of the maximum possible performance. (By adding extra capacity to our system, using more machines or faster machines, we should be able to make that last 10% irrelevant.)

Steps in System Design

On the whole, apart from a discussion of the common technologies available today, the discussion will be as general as possible for the rest of this book, relying purely on objectoriented theory and generic UML notation. One exception to this is that where primitive types must be added to diagrams, Java primitives will be used: this is to avoid confusion when looking at the Java code fragments and because some Java classes have primitive-looking names. (Although this is not a strictly generic use of UML, the UML standard does allow the relaxation.) For most of this book, Java arrays will be avoided in favor of collection classes, since the latter are more elegant.

8.2 DESIGN PRIORITIES Since object-oriented software development is incremental, we mustn’t expect to design the whole system in one pass. So, at the start of each design phase, we need to plan which parts of the system we will design. The use case priorities will help here; so will the use case urgencies that we indicated using the trafﬁc light analogy during the requirements phase: anything green must be fully designed; anything amber doesn’t have to be designed, but it must be supported; anything red must not be designed, but it should still be supported (‘designed’ means that a solution has been invented; ‘supported’ means that a reasonable solution is possible, which requires some foresight on our part). In practice, we seek a system architecture that will support a practical, efﬁcient solution for all the use cases. Within that architecture, we perform detailed design for the most important use cases and partial design for the less important ones. Between increments, we adjust the priorities, the urgencies and the design, as appropriate.

8.3 STEPS IN SYSTEM DESIGN Design can be thought of as having two distinct activities: system design and subsystem design. System design forces us to take a high-level view of the task ahead before we dive into the detail of subsystem design (see Chapter 10). Of course, in good object-oriented tradition, we can blur the boundaries and we can spiral and iterate, but the idea of having two activities is still essential. System design includes the following activities: • Choosing a system topology: how the hardware and processes will be distributed, perhaps over a network. • Making technology choices: selecting programming languages, databases, protocols and so on (see Chapter 9); some decisions may be deferred until later in the design phase.

207

208

Chapter 8

• Designing a concurrency policy: concurrency means many things happening at once – multiple processes, users, machines; these must be coordinated by our software in order to avoid chaos. • Designing a security policy: security has a number of aspects, each of which must be properly addressed and controlled; as an example, consider a customer’s personal data – we must ensure that the data is not stolen by criminals and we must ensure that it can’t accidentally be shown to other customers. • Choosing subsystem partitions: often, it is impractical to produce a single piece of software that solves all of our problems; instead we need to produce separate pieces of software and then make sure that the pieces communicate effectively. • Partitioning the subsystems into layers or other subsystems: typically, each subsystem will need to be decomposed further into manageable chunks before we can do detailed design. • Deciding how machines, subsystems and layers will communicate: communication decisions usually happen as a side-effect of the other steps.

8.4 CHOOSING A NETWORKED SYSTEM TOPOLOGY System topology indicates how a system is decomposed into separate physical and logical components. In this section, we see an overview of the history of network architectures and then discuss current architecture issues: thin versus fat clients; networks; and client–server applications versus distributed applications. We’ll also see how we can use UML deployment diagrams to illustrate architectural decisions.

8.4.1 The History of Network Architectures Most modern, networked, systems have a three-tier architecture. To see what ‘three-tier’ means and to understand why it’s a good idea, we need to take a look at the historical alternatives. Back in the 1940s, computers were large monolithic devices capable of running only one program at a time. These monolithic machines evolved into mainframes that were capable of running multiple programs simultaneously, typically one program per user or one program per batch (a batch comprises sets of similar data that are run through a program in sequence, to process electricity bills, for example). Mainframes were able to handle multiple programs simultaneously because they employed a time-slicer to run each program on the CPU bit-by-bit, each program taking it in turns to perform a little more processing. To begin with, users and batch administrators would access a mainframe via a teletype unit – an electric typewriter that the operator would use to send one line of text to the

Choosing a Networked System Topology

mainframe at a time (the text would represent program commands or program data). The program running on the mainframe would then send back one or more lines as a response. As technology improved, teletypes were replaced with dumb terminals (also known as green screens) that used a Cathode Ray Tube (CRT) instead of paper for textual input and output. With a dumb terminal, the operator could prepare an entire screen of commands or data before dispatching it all at once to the program running on the mainframe (see Figure 8.1).

Dumb Terminals

Mainframe

Figure 8.1: Mainframe computing: a one-tier architecture

The mainframe model, still in use today for large-scale business tasks, is a one-tier architecture. This means that, for any given program, there is only one level of computing activity running on one machine (teletypes and dumb terminals do not perform any processing). Or, to put it another way, one-tier architectures have no network: although there is a thin wire from each terminal to the mainframe, possibly over a phone line, that doesn’t constitute a network in the modern sense. The main advantage of a mainframe is that it’s simple to set up; the main disadvantage is that we can only increase computing power by buying a new mainframe, or by upgrading the existing one. The next architectural generation, popularized in the 1970s, was the two-tier architecture (although it wasn’t called that at the time). The idea was to have processing power on each client, with an optional hard drive, so that large central machines would not have to do all the processing. Thus, we could add, or replace, client machines as a cheaper alternative to changing the central machine. Mini computers and workstations were introduced as clients that accessed midi computers and ﬁle servers, respectively (see Figure 8.2). The mini–midi combination shown here (for example, VT100 and PDP/11) is really the same as the workstation–ﬁle server combination (for example, Solaris workstation and Solaris ﬁle server), they just came out of different communities.

209

210

Chapter 8

Minicomputer

Midicomputer

Workstation

File server

Client Tier

Server Tier

Figure 8.2: Two-tier architectures

With the two-tier architecture, programs and data must be shipped from the large central machine to the client machines; this requires fast connections, leading to the modern idea of a network. A network is an arbitrary number of host machines connected together using fast communication links. (We don’t need to worry about the different network topologies that are possible; sufﬁce to say that a network allows us to get a piece of information quickly from one machine to another.) Once we have a network, we can replicate the large central machines as well as the clients, so it’s much easier to add computing power. Owners of client machines with hard drives also have the ﬂexibility to manage their own data and programs without having to go through the system administrators. If we allow client machines to hang on to programs and data, we have a maintenance headache: as data changes and programs are upgraded, the clients get out of step. It is quite reasonable, therefore, to prohibit the long term storage of programs and data on client hard drives. Two-tier architectures brought sophisticated graphics capabilities and window systems to clients, replacing the text-only model of teletypes and dumb terminals. Two-tier architectures are still widely used today, mostly in the form of Unix workstations and ﬁle servers.

8.4.2 Three-Tier Architecture The three-tier architecture (see Figure 8.3) became popular in the 1990s, as a way of separating user interfaces, program logic and data in networked systems, for reasons that

Choosing a Networked System Topology

Server

Desktop computer Server

Mainframe Client Tier

Middle Tier

Data Tier

Figure 8.3: Three-tier architecture

will shortly become clear. In a three-tier system, any one program involves at least three machines: the data tier stores the data and provides safe concurrent access to it, typically with the help of a database management system (DBMS); the middle tier – also known as the business logic tier or server tier – runs multi-threaded program code using large processors and lots of memory; the client tier presents the user interface to the user, so that they can enter data and view results. The three-tier architecture has many beneﬁts: • Separation of concerns: Any large system has to deal with secure, efﬁcient management of bulk data; high-throughput programming logic; and simple user interfaces. By programming each of these concerns separately, we make the developer’s job easier and we can access the optimizations that have been designed for each particular concern. • Using the right machine for the job: There is no such thing as a ‘one size ﬁts all’ computer. Running a user interface is a simple task that doesn’t require a mainframe or even a ﬁle server; executing programming logic uses CPU and memory intensively but doesn’t require huge amounts of disk space, so a powerful server machine can be used; managing bulk data (for example, a million electricity customers) requires as much computing power and disk space as you can lay your hands on, so this calls for large server machines or mainframes. • Improved performance: We can replicate machines in the data tier or the middle tier, to spread the computing load (load balancing), and each tier is specialized and thus easier to optimize.

211

212

Chapter 8

• Improved security: Frequently, a three-tier system will be deployed in such a way that the client machines are running over the Internet. Therefore, we must have a rigorous security policy in place to protect our internal machines, our programs and our data. With the three-tier architecture, we can make the middle tier secure – impervious to accidental or malicious attack from the outside. Since our data tier is behind the bomb-proof middle tier, we do not have to secure its communications or hardware; this means that our data tier is easier to program and it runs faster. (However, there is a growing feeling that internal tiers need to be protected from employees.) • Protection of our investment: For situations where we have an existing mainframe that has been performing bulk data storage and batch processing for years, we would prefer not to throw everything away and start again, just to get the beneﬁts of a network and a three-tier architecture. (In general, it’s a bad idea to rewrite software: we end up with another ‘Version 1.0’ with the usual lack of features, performance problems and faults.) Instead, as long as our mainframe manufacturer can make the mainframe network-accessible (most have already done so, because of the ‘client–server’ revolution), our middle tier can act as a client to the mainframe and as a server to the real clients. This allows us to transform our legacy system into a modern marvel. • Flexibility: As well as the obvious ﬂexibilities of being able to add and remove machines at will, if we design our system with a three-tier architecture, we get deployment ﬂexibility as well. For example, as long as the logical partitioning is correct, we can develop the system on one tier and then deploy it to three tiers, two tiers (middle and data tiers combined) or one tier (all three tiers combined), as the need arises. • Accommodation of different types of client: Since the client-tier machines only have to capture user input and display system results, we can provide different interfaces for different kinds of device: personal computers, PDAs, set-top boxes, mobile phones, etc. The middle and data tiers can work unchanged in all contexts. Because of the many advantages of the three-tier architecture, it is recommended that you use it for all your systems, even small ones. After all, even the most humble one-tier program is eventually put to the test of ‘Can it be accessed over the network?’ You should only choose a two-tier or one-tier architecture for exceptional cases, such as if you need to optimize application size or speed for special purposes, for example, washing machine controllers or operating systems.

8.4.3 Personal Computers In the 1970s, computer enthusiasts working in their garages began building computers that could be used at home. These self-contained, one-tier machines really took off when endorsed in the early 1980s by IBM – the Personal Computer (PC) design was meant to result in ‘a computer on every desk’.

Choosing a Networked System Topology

By the end of the 1980s, it was obvious that PCs were more useful if they were attached, via a corporate network, to ﬁle servers and mainframes. The result was a two-tier architecture, with PCs performing complex tasks such as spreadsheet editing and document production, while machines on the corporate network provided e-mail and, to a lesser extent, access to corporate data. Since the 1990s, PCs have become increasingly popular. In the corporate sector, they beneﬁted initially from endorsement by IBM, and then from their facilities and their low cost (compared to workstations). PCs have also made home-working more practicable. In the home market, there has really been no competition: the high cost of workstations and the low speed of home access to external networks has meant that there is no realistic alternative. Nowadays, the PC design has displaced all other competitors that emerged from the work of those early garage enthusiasts, with the exception of offerings from Apple (especially popular with graphic designers).

8.4.4 Network Computers By the mid-1990s, as mainframes, Unix boxes and PCs jostled for position in the corporate space, the maintenance cost of PCs was becoming a headache. As mentioned earlier, if you give someone a machine with a hard drive, data and programs on that drive will differ from the central data and programs. This leads to errors – customer data being out of date, for example – and huge extra cost – every client machine has to be returned to central administration for the installation of new releases of e-mail programs, word processors, and so on. As a result, the concept of a network computer was created. A network computer gets all of its data and programs on demand from large central servers or mainframes. This requires a fast network (tens or hundreds of megabits per second). In order to prevent individuals polluting the model, network computers do not have a hard drive. (Well, they can have a hard drive, but only if it’s used by the operating system to cache data and programs.) As you may already have worked out, a network computer is really just a disk-less workstation, typically running some ﬂavor of Unix, because Unix is good at handling on-demand loading of programs over fast networks. Network computers were a very good idea. Not surprising really, since disk-less workstations had been successful for many years. But network computers failed to break the stranglehold of PCs. They even failed to break the stranglehold of workstations with disks, in areas where these were popular. (Even professionals like to keep some autonomy, regardless of the cost to themselves and their employer in frustration and wasted time.) Network computers were used by Java enthusiasts to spread Java in corporations. Since network computers were naturally Unix machines and Java programs can run on Unix, the hope was that Java and Unix would suddenly appear everywhere. However,

213

214

Chapter 8

the network computer has struggled to catch on. Network computers have never been practical for home use because Internet connections, even broadband, are too slow. In order to move beyond PCs, home users need Web browsers that access three-tier architectures. The terms thin client and fat client emerged at about the same time as network computer. The intention was to emphasize the advantages of the lean, mean client machine, typiﬁed by the network computer, over the obese, disk-full PC attached to a network. Thin clients are a good idea: whether you sit in front of a PC or a workstation, all nontrivial work should be network-centric. In the corporate sector, this can be achieved using Java applets accessing a three-tier system. In the home, the best alternative is a Web browser accessing a three-tier system.

8.4.5 The Internet and the World Wide Web By the mid-1980s, many researchers and government employees were making use of a world-wide network of computers, now referred to as the Internet. The Internet owes its origins to US research and defense agencies who had put together ARPAnet decades before. The Internet is characterized by free access to central servers that allow machines and people to locate other machines by their Internet address. Internet addresses have a literal dotted decimal form, e.g. 100.99.88.32, and a symbolic dotted ASCII form, e.g. www.nowherecars.com. Under the covers, the Internet relies on a low-level protocol called TCP/IP: any machine that understands TCP/IP has access to all the facilities of the Internet. In the early 1990s, when the Internet was ﬁrmly established in the research community (as a tool for e-mail and ﬁle transfer), Tim Berners-Lee, working at the Swiss research institute CERN, came up with the idea of documents that contain hyperlinks to other documents on the Internet. Berners-Lee made his ideas practical by inventing a document layout language called HyperText Markup Language (HTML) and a protocol called HyperText Transfer Protocol (HTTP). HTTP allows any machine to load a document via a hyperlink from any other machine. The location of a document can be speciﬁed using a Uniform Resource Indicator (URI) of the form http://www.nowherecars.com/index.html. Berners-Lee’s invention was christened the World Wide Web (of information). By the mid-1990s, anyone wanting to make documents available for others to read could deploy them on a Web server and anyone wanting to read documents could run a Web browser on their client. Once Internet Service Providers appeared, providing access to the Internet from home, the World Wide Web spread everywhere. Nowadays, the distinction between the Internet and the World Wide Web is blurred. You may hear any number of terms used more or less interchangeably, such as Internet, Net, World Wide Web, Web or Information Super-Highway.

Choosing a Networked System Topology

8.4.6 Intranets The Internet has two major problems: it is slow and it is insecure. The performance problems come from the fact that information might have to travel thousands of kilometers before it reaches its destination, often encountering slow connections and servers on the way. Also, the information has to compete for delivery with all the other information that’s in transit. The insecurity of the Internet results from the fact that everyone has access and, therefore, anyone can intercept and read information in transit – even if we encrypt the information, there is still potential for someone to decrypt it. As a result, the concept of an intranet was invented, to mean a ‘mini Internet’ running behind closed doors. Intranet means ‘network inside’, as opposed to Internet, which means ‘network between’. An intranet is typically controlled by a single corporation or government. By preventing access to an intranet from outside, we can improve performance (using powerful machines with no competition from outside trafﬁc) and we don’t need to worry so much about security. For global intranets, information may still travel long distances, so, on the whole, we still have to forego the performance of a local area network. As far as security is concerned, we need to make sure that no unauthorized person comes to one of our sites and connects their own computer, of course; beyond that, we have to deal with all the usual problems of industrial espionage, disgruntled or careless employees, spies monitoring the electro-magnetic emissions from our machines and networks – rumor has it that one government organization makes its employees use their computers inside Faraday cages, supplying mains power by induction from outside the cage: it’s all a question of how paranoid you want to be. For most intranets, we can retain the local performance and security beneﬁts and still give our employees access to the Internet (for e-mail, information gathering and monitoring competitors’ Web sites). To do this, we need to employ an Internet ﬁrewall. A ﬁrewall is a piece of software that allows machines on an intranet to access any TCP/IP address, while making intranet addresses invisible from the outside. Internet ﬁrewalls can perform other tasks too: for example, we can make sure that only Web browsers can access the Internet from employees’ machines – this prevents rogue programs from sneaking in and communicating secrets to the outside world; we can conﬁgure our Internet ﬁrewall to allow e-commerce requests from outside to tunnel through on their way to our e-commerce servers; at home, we can use a personal ﬁrewall to protect our PCs from hackers and viruses.

8.4.7 Extranets and Virtual Private Networks So how can we exploit the performance and security beneﬁts of intranets for business-tobusiness communications? Easy – we use an extranet, a secure connection between one or more intranets. The term extranet is a pun on intranet, short for ‘external intranet’. The

215

216

Chapter 8

easiest way to make an extranet is to run a piece of software on the Internet ﬁrewalls at the edge of each intranet, which does two things: • Allows information to pass from ﬁrewall to ﬁrewall. • Uses strong encryption to protect that information as it passes over the Internet. An extranet is also referred to as a Virtual Private Network (VPN). The simplest form of VPN, even older than the idea of an extranet, is a worker dialing in to their corporate LAN from home using purpose-built software. Figure 8.4 illustrates how intranets, extranets and the Internet ﬁt together. Intranet

Intranet Extranet

Firewall

Firewall

The Internet

Figure 8.4: How the different kinds of net ﬁt together

8.4.8 Client–Server versus Distributed Architectures Whenever we connect multiple machines or software systems, we have to choose between client–server and distributed styles, as illustrated in Figure 8.5. Despite its origins in the mainframe arena, these days ‘client–server’ simply means that we have a large number of

Choosing a Networked System Topology

Client – Server Architecture

Distributed Architecture

Figure 8.5: Client–server and distributed architectures

small, simple clients sending requests to a few large, multi-threaded servers that process the requests. A distributed (or peer-to-peer) architecture, on the other hand, is characterized by a collection of autonomous peers, communicating in any direction, as the need arises. The most common example of a client–server architecture is the e-commerce model: customers’ Web browsers issue requests to company Web servers, which in turn issue instructions to back-end systems. Most two-tier and three-tier systems are client–server. A good example of a distributed architecture is when a massive computing task is spread over many Internet machines. If we have a very large amount of data or processing that needs to be performed and we can partition the data or the processing, we can distribute the problem over independent machines. One example is SETI@home, a nonproﬁt organization that looks for alien radio signals. The search for extra-terrestrial intelligence (SETI) started out as a NASA project with radio telescopes scanning the heavens and recording anything they came across. This data was then processed to see if it contained evidence of coherent radio transmissions from aliens. Eventually, the SETI project was canceled (because of lack of funds or lack of belief) and was reborn as SETI@home. The enthusiasts who developed SETI@home take the same radio data and distribute it to machines on the Internet: each participating machine analyzes its data using a piece of software that runs as a screensaver; any interesting results are sent back to the central server for further analysis. For more information, see the SETI@home Web site at setiathome.ssl.berkeley.edu.

217

218

Chapter 8

Although SETI@home has a central repository of data, it is still a distributed architecture because most of the processing is carried out by independent nodes. The idea of bringing together huge numbers of machines to solve a complex task has also been used for primenumber searches and cancer research. The basic idea is now being researched under the heading of grid computing. The terms ‘client–server’ and ‘distributed’ (or ‘peer-to-peer’) are also used to describe software architectures, independently of how the software is deployed on physical machines and networks. Objects running in a program are a good example: normally, we write our objects as servers that can be reused in different contexts with different client objects; but, for special purpose applications, we might also write groups of objects that collaborate in a distributed manner. Network communication links tend to be bi-directional, meaning that although the link may be opened initially by the client, the server can also send information to the client (whether the client takes any notice is up to the designer of the client software). Thus, strictly speaking, the distinction between client–server and distributed is an artiﬁcial one, used by designers to structure their solutions in one way or another. As a general rule, client–server architectures are easier to develop, but they may not provide maximum theoretical performance (for example, the client is usually idle when the server is processing one of its requests). Distributed architectures are, usually, harder to develop but they may provide better performance. Often, the choice of architecture is obvious. For example, making a purchase is a step-by-step process (customer asks for product details, supplier provides product details, customer elects to purchase product, supplier provides purchase form, customer submits purchase form, and so on), and thus is a natural client–server interaction. In contrast, a multiuser ﬂight simulator requires each pilot’s computer to perform complex real-time rendering of the cockpit, scenery and other aircraft – the only communication needed is the periodical broadcast by each machine of its plane’s current position. Thus, a multiuser ﬂight simulator is a natural distributed architecture.

8.4.9 Depicting Network Topology in UML System architectures can be depicted in UML on a deployment diagram (see Figure 8.6). This simple deployment diagram shows only nodes, communication paths and multiplicities. Each node, in this diagram, represents a host machine (indicated with the UML keyword <<device>>). A communication path shows that two nodes communicate in some way. Nodes can be given multiplicities to indicate how many might exist at run time: thus, in this diagram, we have nodes that are duplicated (CootServer and DBServer) and nodes that are multiplied (CootHTMLClient and CootGUIClient). Deployment diagrams are similar to class diagrams and object diagrams in that they can show possible architectures (node types) and actual architectures (node instances). When

Choosing a Networked System Topology

multiplicity

* <<device>> CootHTML Client

* <<device>> CootGUI Client

2

node

communication path

<<device>> Coot Server

2

<<device>> DB Server

Figure 8.6: A basic deployment diagram for Coot

showing node instances, just like objects on an object diagram, the node label takes the form name:Type, and should be underlined. Most deployment diagrams need an accompanying description if they’re to make any sense. We might call this a deployment survey, in Jacobson style.

Case Study iCoot deployment survey This description could accompany Figure 8.6: The iCoot data tier comprises two database servers (which we have called Having two such nodes improves throughput and reliability.

DBServer).

219

220

Chapter 8

Case Study (cont’d) The middle tier, which communicates with the data tier, consists of two server machines (CootServer), again duplicated for the sake of reliability and throughput. Each CootServer can be accessed simultaneously by any number of CootHTMLClient nodes. Eventually, we will also provide access from CootGUIClient nodes.

8.5 DESIGNING FOR CONCURRENCY Most systems, especially networked systems, have many things happening at once; that is, they are concurrent systems. This has implications for the system as a whole and also for the individual processes running as part of the system. Although it would be much easier to develop systems if we could rely on all users and processes to form an orderly queue, in reality we have to turn chaos into order with our own programming efforts. Concurrency introduces the same issues, time and again: • How to ensure that information is updated completely before anyone can act on the update; for example, to stop anyone accessing the details of a new car model until all the details have been added. • How to ensure that information is not updated while it’s being read; for example, don’t delete a car model while its details are being viewed. At a low level, database transactions and thread monitors are used to protect data inside individual processes, for example. At a higher level, we need to use system rules and business rules to control concurrent activity. The easiest approach to concurrency is to constrain the system or introduce extra business rules, especially if the user experience doesn’t diminish signiﬁcantly. For example, for iCoot, rather than trying to cope with updating the car catalog while customers are accessing it, we could update the catalog in a separate database and switch databases once a day; this way the Internet subsystem can assume that the catalog is read-only, which makes our code much easier to write. (The user experience diminishes slightly – in rare cases they will have to be told that a car model they’ve just tried to reserve has been discontinued – but we choose to live with that.) This is an artiﬁcial constraint on the system.

Designing for Concurrency

Business rules can also make the developer’s life easier. Indeed, business rules are sometimes introduced because there is simply no off-the-shelf way to solve the problem. For example, consider a purchasing system for concert tickets. Customer Fred walks into a booking ofﬁce in Paris at about the same time that customer Beryl walks into another booking ofﬁce in New York. Both customers have decided that they want a ticket to the same concert in London. Unfortunately, there is only one ticket left. How do we decide who gets the ticket? Both customers will ask ‘Is there a ticket available?’ Both sales clerks will check on the system and both will reply ‘Yes’. Now we have a race: the ﬁrst customer who says ‘Okay, I’ll take it.’ will win, subject to the efﬁciency of the clerks and the network delays from Paris and New York to the location of the actual server. In the ticket scenario, we must ensure, at the very least, that we don’t inadvertently sell two tickets when there’s only one seat left. This is a process-level concurrency issue, because it can be controlled by the server process (the requests are serialized: the ﬁrst to arrive gets the ticket, the second gets an error message). But that may not be good enough from the business point of view, because we end up with one disgruntled customer who was told there was a ticket only to be told a few seconds later that there wasn’t one after all. To avoid disgruntled customers, we can introduce an extra business rule: when a clerk queries the availability of a ticket, if one is available, it is temporarily reserved – the reservation lasts until the clerk cancels the enquiry or the reservation times out (for example, the temporary reservation lasts for ten minutes if the clerk doesn’t cancel it ﬁrst). With this new business rule, we can ensure that only the ﬁrst customer who makes the virtual trip to the ticket server is told that there’s a ticket available (the ticket server can combine the query and the reservation into a single business service). Let’s assume that Fred is successful. The clerk serving Fred has ten minutes to persuade him to pay for the reserved ticket, or the clerk can cancel within that time if Fred changes his mind. (The clerk would also cancel if Fred’s method of payment failed, which is another reason for having temporary reservations in the ﬁrst place.) Beryl now thinks that the concert sold out before she got to the booking ofﬁce, so she has no reason to blame the booking agency for providing incorrect information. (If the last ticket is not bought by Fred and Beryl later discovers that the concert was not sold out, she can simply be told ‘Somebody must have canceled’.) Detailed consideration of concurrency problems and their solution is beyond the scope of this book. In any case, there is no real substitute for sitting down and thinking (hard) about it. For now, here are some observations: • The look and feel of a well-designed concurrent system is no different to the single-user version. • Our business services are the same for concurrent and single-user cases.

221

222

Chapter 8

• To make a business object concurrent-safe, it’s only necessary to add messages and supporting objects; therefore, business messages (and associated attributes) can be designed separately. If you can think of a concurrent situation that might cause difﬁculties for your system, do not proceed to implementation until you can guarantee that that situation is no longer a problem. Be tough on your system, because reality will be tougher.

8.6 DESIGNING FOR SECURITY Detailed consideration of security, although fascinating, would require a book in itself. Therefore, only an overview will be provided here. A secure system is one that is protected from misuse, regardless of whether the misuse is accidental or malicious. Security is a rather broad term that can be broken down into the following ﬁve aspects: • Privacy: We must be able to hide information, making it available only to those who are authorized to read it (or change it). • Authentication: We need to know where each piece of information came from, so that we can decide whether or not to trust it. • Irrefutability: This is the ﬂip-side of authentication, ensuring that the originator of information can’t deny that they’re the source; this is helpful to us if anything goes wrong. • Integrity: We must be sure that information hasn’t been damaged, accidentally or maliciously, on its way from the source to us. • Safety: We must be able to control access to resources (such as machines, processes, databases and ﬁles). Safety is also known as authorization. In this context, information means not only data, such as business documents and user passwords, but also pieces of executable code. Code is an issue because it may be loaded dynamically over the network. The ﬁrst four requirements above can be satisﬁed using digital encryption (more on this in the next section). The safety requirement is rather more complex. Normally, when a piece of code is running, the operating system will exert some kind of control over what the code is able to do; typically, this means controlling access to ﬁles, to directories and to other programs. Operating systems have faults and security holes (some more than others). In addition, the control provided by operating systems is inﬂexible: for example, we have no easy way for the user or the developer to invent their own abstract controls such as ‘Only Freda and Ben can join my virtual game of poker’ or ‘Only members of my family can turn the microwave on using their mobile phone’.

Designing for Security

If our system operates over a network, the safety aspect becomes even more important. This is because, in a networked world, hackers can try to hijack programs running on our machines to get them to misbehave. To make matters worse, courtesy of Java and ActiveX, code fragments can travel around the network, executing on different machines as they go. Java is the only mainstream technology that adds the required level of safety for code traveling around a network and the ﬂexibility to invent our own controls, regardless of the operating system being used. The practical implication of this is that if you’re not using Java, you should prevent code from moving around the network: install code on each client or server as needed and rely on the operating system and your own programming efforts to make things safe.

8.6.1 Digital Encryption and Decryption This is an aside to give you some idea how digital encryption and decryption – cryptography – can be used to provide privacy, authentication, irrefutability and integrity. First, the basic idea: to encrypt information means to scramble it so that it’s useless if anyone manages to steal it. For this to make sense, the scrambling method must be reversible and it must be known to the intended recipient. The reverse of encryption is decryption. As kids, most of us played with the so-called Caesarian Cipher (although we probably didn’t call it that). What we do is to take a message and then scramble it by shifting each letter a number of places down the alphabet, wrapping round if we fall off the end. For example, if we choose to shift four places, the message ‘Mister Watson, can you hear me?’ would be encrypted as ‘QMWXIVAEXWSRGERCSYLIEVQI’. (We leave out punctuation, letter case and spaces in the encrypted version in order to avoid giving clues to spies.) As long as the recipient knows that the algorithm is Caesarian and the offset is four, they can decrypt the original message. The Caesarian Cipher is a general form of encryption: it involves an extra piece of information that the recipient must know (4) as well as the algorithm itself (Caesarian). The extra piece of information is called a key, because it unlocks the code: it’s not enough to know that unlocking a door involves turning a key, you need to have the key too. The success of an encryption mechanism boils down to two things: • How difﬁcult it is to crack the encryption, by trial and error or by hard work. • Safe distribution of the key to the intended recipient. Both of these can have their problems. The Caesarian Cipher, for example, is easy to crack: a wrong-doer only has to guess that the Caesarian method is being used and then try a maximum of 26 possible offsets before the message materializes in front of their eyes. As for safe distribution, we might try whispering the details of the key into the ear of every intended recipient: however, as well as being inconvenient, this approach is doomed to failure – if any

223

224

Chapter 8

recipient with a defective memory writes the key down on a piece of paper, we might have a security breach. Although it’s possible to come up with an uncrackable encryption algorithm – one-time pads and quantum cryptography are examples – these algorithms are either inconvenient (code books) or very expensive (quantum cryptography, at least for the time being). So, for everyday use, especially in the context of network programming, we normally choose the level of security that we want and then select an efﬁcient algorithm that provides that level. This works because no self-respecting criminal would bother to eavesdrop our encrypted credit card details if it might take them, say, a million years to decipher the details: they would just burgle our house or pick our pocket instead. Digital encryption and digital decryption are simply an extension of the ideas above to encrypt the digital information passed between computers, where software can do all the hard work. The level of digital encryption is normally expressed in bit strength: 128-bit encryption is considered minimum these days and 1024-bit encryption is desirable. Digital keys are based on prime numbers, hence cracking them is difﬁcult, because it involves trying to ﬁnd prime factors, which is an intensive process. The keys themselves are distributed using digital certiﬁcates. A certiﬁcate, with the help of a trusted certiﬁcate authority, authenticates the key, so that we can’t be fooled by hackers. Below is an indication of how four of the security aspects identiﬁed earlier can be implemented using cryptography: • Privacy: In digital encryption and decryption, safe distribution of keys is achieved using public/private key pairs, certiﬁcates and certiﬁcate authorities. • Authentication: This relies on being able to prove the origin of the key, using certiﬁcates and certiﬁcate authorities; roughly speaking, if we can successfully decrypt the information, and we know the origin of the key, we know that the information must have come from the same place as the key. • Irrefutability: Since authentication relies on proving the origin of the key, once we’ve authenticated a piece of information, our authentication is irrefutable. • Integrity: First, we transmit the encrypted information and the unencrypted information to the client. Next, the client decrypts the encrypted version and compares the result with the unencrypted version – obviously, the two should match (the chance of them matching by accident is tiny). Therefore, we’re conﬁdent that we have received the correct information. There are two problems with integrity-checking as just described: ﬁrst, we’ve sent the information unencrypted; second, we’ve sent the information twice (the encrypted version is just as big as the unencrypted version). The ﬁrst problem can be overcome by sending

Partitioning Software

everything over a secure line. The second problem can be overcome by an optimization based on message digests – a digest is a small string of bits generated from a piece of information using an irreversible algorithm. Without going into detail, we end up sending the encrypted digest, but not the encrypted information. The end result is that we get the same degree of integrity checking, but the information is sent only once. (For the record, an encrypted digest is called a digital signature.)

8.6.2 General Security Rules As with concurrency, when designing a secure system, there is no way to avoid sitting down and thinking hard about it in order to make sure that no-one can abuse the system. Here are a few things to bear in mind when trying to secure your (networked) system: • Protect your servers from unauthorized access, whether accidental or malicious. • Conﬁne sensitive information to your internal network: sensitive information includes details of business deals with other companies; business strategy; personnel details; details of the credit reference agencies you use; information relating to national security; and so on. • Prevent the eavesdropping of exported information: ensure that information you pass outside your intranet can only be read by the intended recipient. • Protect employee and customer passwords, which are not only the foundation of your entire security policy, they’re often highly personal. • Prevent server code accessing unneeded resources. • Prevent client code accessing unneeded resources: we want to protect the client against unauthorized access to their resources and against accidental damage (because we want to offer a high quality of service and because we don’t want them to sue us if something goes wrong). If you’re really brave, you might consider hiring ethical hackers, consultants who do everything they can to break into your system. Once the ethical hackers have certiﬁed your system, you’re less likely to get a nasty surprise when it goes live.

8.7 PARTITIONING SOFTWARE For any large business, it’s impractical to lump all the business entities and business processes into a single software system – the result would be far too complex and difﬁcult to use. We can, and should, partition software into systems, then into autonomous subsystems if necessary, then into layers (which can also be considered to be subsystems).

225

226

Chapter 8

8.7.1 Systems and Subsystems Consider how the simple concept of Customer is viewed differently by the departments in a large organization, sales, marketing, billing, procurement, dispatching and so on. If we tried to put together a single software system that supported all of these departments, our Customer would have hundreds of attributes and hundreds of operations: a recipe for disaster. Instead, a business should have a number of separate systems, each implemented by a different development team so that the temptation to reuse objects inappropriately is minimized. Then, where information needs to be passed between systems, it should pass in a well-deﬁned, controlled manner via well-deﬁned interfaces. In order to reduce complexity further, each system should be broken down further into separate subsystems. Figure 8.7 shows a company’s systems as independent trees in a forest. Underneath each tree is the database of information that the systems need to access and at the top is the user interface. Communication takes place along narrow, constrained pathways (rope bridges), from business logic to business logic via purpose-built interfaces. Although each system has its own independent data, we can still use a single DBMS for deployment, since a DBMS will happily manage multiple databases.

Figure 8.7: Coordinating multiple systems

For the case study, we have a single system called Coot which comprises two subsystems: iCoot provides access for members and nonmembers; the other subsystem provides access for assistants and resembles the existing Auk interface. The latter subsystem will not be considered in detail.

Partitioning Software

8.7.2 Layers Inside a software system, we usually employ multiple layers of code (see Figure 8.8). Each layer is a cluster of collaborating objects dependent on the facilities offered by lower layers. Layers don’t have to contain objects. For example, the Unix system library provides access to low-level operating system facilities via a layer of C functions.

Figure 8.8: Objects in layers

Layers help to reduce complexity by breaking the implementation up into more manageable chunks. Layers also increase the chance of reuse, since each layer is written to be independent of the layers above. As Bertrand Meyer put it in [Meyer 97]: A serious software system, even a small one by today’s standards, touches on so many areas that it would be impossible to guarantee its correctness by dealing with all components and properties on a single level. Instead, a layered approach is necessary, each layer relying on lower ones. Often, regardless of the total number of layers, the top layer represents the user interface and the bottom layer represents the operating system, or a network connection. In the following diagrams, the lowest layers have been omitted for simplicity, stopping at a wellknown layer that’s documented elsewhere. In the following sections, we’ll see examples of common layers used in systems that have one, two or three tiers. Layers may be open (exposing some of the objects from lower layers for use by layers above, i.e. it manages the objects below but doesn’t completely hide them) or closed (completely encapsulating the layers below, i.e. objects from lower layers are hidden from those above). There is no easy rule for deciding whether a particular layer should be open or closed. You have to use your skill, judgment, experience and foresight. Generally speaking, a closed layer requires more coding and may run more slowly than an open layer (because

227

228

Chapter 8

there’s more copying and translating of information to be done). On the other hand, an open layer is generally not as safe (because lower layers are unprotected) and harder to maintain (because each layer has more than one dependent layer above it). To a certain extent, we can swap layers in and out without disturbing other code. For example, we can always throw away the top layer and replace it with another. Furthermore, we can replace a closed intermediate layer with one that has the same interface, without affecting the layers above. Most often, layers are re-implemented at the bottom, when we move our system to another platform, or at the top, when we move the user interface to a new device. Some of the technologies in the following discussion are explained further in Chapter 9.

Layers for Single-Tier Systems Figure 8.9 shows a simple layering scheme as might be employed in a single-tier system. At the bottom, we have the database layer, whose job it is to ship data back and forth between the DBMS and the business layer. The assumption has been made that most applications have data storage requirements, so that data doesn’t evaporate if the system is shut down for any reason. If we had a simpler system that stored data in ﬁles, the database layer would be a ﬁle system instead of a DBMS. On top of the database layer, we have the business layer, which consists of the entity objects and supporting implementation objects. Finally, on top of the business layer, we have the user interface layer, which contains objects whose job it is to present available options to the user, to pass user commands and data on to the business layer and to display data coming back from the business layer.

User Interface Business Database

Figure 8.9: Layers in a single-tier system

Layers for Two- and Three-Tier Systems Two-tier and three-tier systems employ a network to get from the user interface running on the client to the business layer running on the server. We can cope with this by employing two more layers, as shown in Figure 8.10. The network layer contains objects that make the network transparent to the user interface. As far as the user interface is concerned, it may just as well be accessing the server objects directly. The server layer contains objects that simplify use of the business layer into a manageable set of business services. As well as

Partitioning Software

User Interface Network Server Business Database

Figure 8.10: Layers in a two tier or three-tier system

simplifying the client, this allows us to localize security measures in the server layer without polluting the business layer and to accommodate different kinds of client. For a two-tier system, the database layer is on the same machine as the server and business layers. For a three-tier system, the database layer spans a network, but the details are hidden from us by the DBMS. If we use HTML forms to get from the client to the middle (or server) tier, the user interface–network situation is less clear cut. In this case, the user interface is partly on the client (as HTML pages and forms) and partly on the server (as servlets and JSPs). If we prefer, we can adjust the diagram to reﬂect this.

Translation Layers As illustrated in Figure 8.11, different layers have different focus. For example, when designing a user interface, we’re concerned with menus, dialogs, notebooks, windows, usability, intuitiveness, and so on. For the network, we’re worried about protocols, bandwidth and different types of hosts. When we move on to the server, we should concern ourselves with security, multi-threading and throughput. In the business layer, the part that ﬂows most directly from our business analysis, we’re most interested in abstraction, attributes, operations, polymorphism, reuse and the other fundamentals of object-oriented modeling. Finally, we arrive at the database layer where, traditionally, we expect to deal with keys, tables, SQL, locking, functional dependencies and all the other aspects of database theory. In short, if we try to connect these differing worlds together directly, we will end up with too much complexity and too much coupling (strong coupling, when one object’s implementation is closely dependent on another’s, makes code harder to maintain). We can reduce complexity and coupling by introducing extra layers to act as translators. Translation layers are particularly useful for translating the business layer (in the single-tier case) or the network layer (when we have multiple tiers) into the minimal

229

230

Chapter 8

Figure 8.11: Differing concerns in a large system

functionality needed by the end user – such layers are often referred to as controllers. Controllers manage the user interface’s communication with the rest of the system. (This ﬁts well with Jacobson’s idea of a controller.) Another common translation layer is the so-called persistence layer that sits between the business layer and the database layer: by removing the business layer’s dependence on the actual storage mechanism in use, it becomes easier to change the storage mechanism later (from ﬁles to a DBMS, for example). Figure 8.12 shows a multi-tier system with the addition of control and persistence layers.

8.7.3 Java Layers: Applet plus RMI In order to illustrate a Java client with a proper GUI, Figure 8.13 shows a complete set of layers for a three-tier system with an RMI applet as the front end. RMI is a Java network protocol. In this illustration, the user interface layer is implemented using the Swing library (Java’s portable library of GUI components). Underneath the user interface layer, we have a control layer that contains all the code for accessing the business services; this code would otherwise

Partitioning Software

User Interface Control Network Server Business Persistence Database

Figure 8.12: Translation layers in a multitier system

Swing Control RMI Server Business Persistence JDBC

Figure 8.13: RMI applet layers

have to be buried in the user interface objects and re-implemented for every new interface that we add later (mobile phones, for example). 95% of the network layer is provided for us by the RMI framework – we just have to follow a few simple rules in the control layer and the server layer to make our server objects accessible from any client. The server layer, business layer and persistence layer are just like the ones described in the previous subsection. Last, but not least, we have the database layer provided by the Java DataBase Connectivity (JDBC) library. JDBC allows us to access any relational database using dynamic or precompiled SQL. Because we have included a persistence layer, we could

231

232

Chapter 8

replace JDBC with an object-oriented database, or the ﬁle system, without disturbing the business objects.

Case Study iCoot layers As an illustration of the HTML/CGI-with-servlets conﬁguration, Figure 8.14 shows the layers that are used for iCoot. For iCoot, there is no persistence layer, because a relational database, accessed via JDBC, is expected to serve the purposes of the system throughout its life. The JDBC layer itself is provided by classes from the standard Java library. For the user interface, there are two styles: CGI-with-servlets and RMI-plus-applets. For the ﬁrst increment, HTML/CGI-with-servlets will be used (using JSPs for the dynamic page content) in the ServletsLayer. For later releases, an RMI-plus-applets mechanism will be provided, for desktops, using the SwingLayer, and then small devices, such as PDAs and mobile phones, using the MicroLayer. The MicroLayer is implemented using the Java 2 Micro Edition (J2ME). For the sake of the ﬁrst increment, the control layer is provided by servlets, with the help of JSPs. For the RMI version, on the other hand, the ControlLayer is designed for a proper GUI. Oddly enough, this means that the control layer is below the network for HTTP/CGI, but above the network for RMI. In either case, the controllers are translating (or coordinating) the ServerLayer. Thus, everything from the ServerLayer down is reused without change. (The ControlLayer is also reused, by the two styles of GUI.)

HTMLLayer

SwingLayer

MicroLayer

HTTPCGILayer

ControlLayer

ServletsLayer

RMILayer

ServerLayer BusinessLayer

JDBCLayer

Figure 8.14: Layers for iCoot

Partitioning Software

8.7.4 Message Flow in Layers In a layered system, each layer is a client of the layer immediately below it. Thus, we expect messages to be sent from the upper layer to the lower layer, as shown in Figure 8.15. Each message is either a question (retrieving some information, for example, getAddress) or a command (an instruction to do something, for example, setAddress).

Figure 8.15: Messages ﬂow downwards in a layered system

Many commands sent into a layer will have an effect on the information managed by that layer – otherwise there would be little point in issuing the commands in the ﬁrst place. But, what if the upper layer needs to know what information has changed? The upper layer might, for example, be a user interface that needs to update its display with the new information. We have two choices: • Add knowledge to the upper layer about which commands change which information. • Get the lower layer to send messages to the upper layer whenever the information changes. The problem with the ﬁrst option is that it pollutes the upper layer with knowledge that, logically, belongs in the lower layer – this makes the coding of the upper layer more complicated and couples the upper layer more tightly to the lower layer. The problem with the second option is that the lower layer has to know something about the upper layer so that it knows which objects to send messages to – thus, the lower layer is polluted with knowledge about the upper layer, making it more complicated and also coupling it to the upper layer. (Ideally, lower layers should not be coupled to upper layers at all.)

Events So, is there some way that a layer can notify the layer above when something interesting has happened, without increasing complexity or coupling in either direction? The answer is ‘Yes, events’. Figure 8.16 shows the real-world analogy that we use for events. An event source detects when something interesting has happened (an event) and shouts out the details (broadcasts) to anyone who might be listening (the event listeners). An event might be an attribute event, indicating a change in the value of one of the event source’s attributes,

233

234

Chapter 8

Event Listener

Event

Event Source

Figure 8.16: The event analogy

or it might be a pure event, not related to any attribute value. For example, consider a Clock entity: it could broadcast an attribute event when the time changes – say, every second – and it could broadcast a pure event when the alarm goes off. As the phrase ‘broadcast an event’ suggests, we would like the event source to detect when the event happens and just shout it out to anybody who is listening: this way, the event source won’t be complicated by knowledge of the listeners and the event listeners won’t be complicated by knowledge of when the events might occur. In our layering scheme, we can use event sources in each layer to broadcast events to listeners in the layer above, thus achieving our goals of keeping knowledge in the right place and minimizing coupling. But how do we implement such an idea using messages? As a ﬁrst step to understanding how this might work, Figure 8.17 shows an example collaboration between a Clock, the event source, and a ClockUI, the event listener. In this diagram, the special name self is UML notation for ‘the current object’ (it corresponds exactly to the special name this in Java). As part of the initialization process, the ClockUI registers itself with the Clock by sending it the addClockListener message. Next, the ClockWatcher sets the alarm in the ClockUI, which passes the alarm setting on to the Clock. As the clock sends itself the tick message periodically, eventually it will detect that it is time for the alarm to go off. When this happens, the Clock creates a ClockEvent (with the help of the ClockEventHome), with information about the event (in this case, the only information is the source of the event). The Clock then sends the alarm message to the ClockUI with the ClockEvent as a parameter. (The event object records the

Partitioning Software

2b:setAlarm(time) 1a :a

3c.3:beepBeepBeep()

dd C f)

:a

e)

.2

m

el

(ti

3c

(s

m

er

ar

en

Al

st

et

Li

:s

ck

.1

lo

2b

:ClockUI

Clock Watcher

la rm (e )

:ClockEvent

() ck :ti 3c

3c

.1

:e

=c

re

at

e(

se

lf)

:Clock

:ClockEvent Home

Figure 8.17: A communication diagram showing how events work

event source; if the ClockUI is listening to more than one clock, it can ﬁnd out which one has just sounded the alarm.) Finally, the ClockUI beeps at the ClockWatcher. You may be wondering why the sequence numbers in this communication diagram have letters in them. The reason is that we have three independent sequences in our scenario: adding the clock listener; setting the time; and sounding the alarm. (As long as adding the clock listener happens ﬁrst, the other two sequences can happen multiple times, in any order.) The clock has to run independently of the ClockUI and the ClockWatcher, otherwise, when the alarm has been set, control returns to the ClockWatcher and the Clock is not able to tick. A name as part of a UML sequence number shows that the message in question is dependent on messages that include the same name, but independent of others. We can still show the ordering of independent messages, by the careful use of numbers: for example, it is implicit that message 2b happens before message 3c in our scenario; it would also be implicit that messages numbered 99x and 99y happen at the same time.

235

236

Chapter 8

The class diagram for the alarm clock scenario is shown in Figure 8.18 (for simplicity, and ClockWatcher have been omitted). From this diagram, we can see that ClockEvent has a getter for the source attribute. Also, the ClockUI has a message that allows the ClockWatcher to set the alarm – setAlarm – and another message for detecting the alarm event – alarm. Finally, the Clock class has a message for setting the alarm – setAlarm and a message for registering a listener – addClockListener. (There are some new pieces of UML notation in this diagram which will be explained shortly.) ClockEventHome

<> ClockListener

ClockEvent source:Clock

alarm(:ClockEvent)

getSource():Clock

realization

listener

ClockUI

source

*

alarm(:ClockEvent) setAlarm(:int)

Clock

* addClockListener(:ClockListener) setAlarm(:int) ClockListener

provided interface/required interface

Figure 8.18: Events as a class diagram

Because object-oriented languages generally don’t have a true broadcast mechanism, the Clock in Figure 8.18 must maintain an internal list of interested listeners and then send a message to each one when the event occurs. The listeners, for their part, must make sure that they register for the event. It may have occurred to you that the Clock is now coupled to the ClockUI, something that we were trying to avoid in the ﬁrst place. We can solve this problem by introducing an abstract class – called ClockListener – that only lists the messages required for detecting Clock events. As long as ClockUI inherits from ClockListener, we will be able to register a ClockUI with a Clock and the ClockUI will be able to receive the alarm message. Thus, although Clock is coupled to ClockListener, it is not coupled to ClockUI. (ClockListener resides in the same layer as Clock, while ClockUI ﬂoats independently in the layer above.) An interface, denoted by the <> keyword, is a pure abstract class – a class that has no concrete methods and no attributes. Since interfaces are so useful, for specifying communication with reduced coupling, they have some special UML notation shown in Figure 8.18. The dashed arrow with a white head (labeled realization) indicates inheritance, for the special case where the superclass is an interface. The ClockListener notation, labeled

Summary

provided interface/required interface,

allows us to indicate that one class uses another class via a particular interface. In this case, the part that looks like a lollipop indicates that ClockUI realizes (is a kind of) ClockListener, while the cup-shaped part indicates that Clock only depends on the fact that ClockUI is a ClockListener. As you’ve probably spotted, there is redundancy in Figure 8.18. It’s up to you when you use the ‘dashed inheritance’ or ‘lollipop-and-cup’ styles in your own diagrams.

Message Flow using Events Figure 8.19 shows how messages should ﬂow when using layers. Ordinary messages are shown ﬂowing down through the layers, while event messages are shown rising up. (The event messages have been shown as dashed arrows to indicate that the lower object has no knowledge of the recipient.)

Figure 8.19: Control ﬂow using events

Events are most often used in client code, because they’re a convenient way of keeping a user interface up to date with the information in the layer below. Events are rarely used on the server side, because server code tends to be multi-threaded, which complicates eventdriven programming signiﬁcantly (there is increased potential for deadlock, for example), and because events should not be broadcast over a network: we don’t want the server held up while it sends event messages to lots of clients, some of which might have perished or be difﬁcult to contact. Therefore, it’s recommended that you avoid event-driven programming on the server side. If you wish to broadcast over a network, use machine-to-machine messaging instead, such as that provided by the Java Messaging Service (JMS).

8.8 SUMMARY In this chapter, we’ve looked at: • The steps involved in system design and how a system can be decomposed into physical and logical components, with particular attention paid to network topologies.

237

238

Chapter 8

• How to illustrate architectural decisions on UML deployment diagrams. • The concurrency issues that arise in networked systems: how to ensure that information is updated completely before anyone can act on it and how to ensure that information is not updated while it’s being read. • How to ensure that a system is protected from accidental and malicious misuse, by ensuring privacy, authentication, irrefutability, integrity and authorization. • How a company’s software can be broken down into multiple systems, subsystems and layers.

FURTHER READING Full details of the HTML speciﬁcation can be found in [W3C 99]. For a more readable description, see [Raggett et al. 97]. Raggett is one of the main forces behind the UML standard. For more information about the history of cryptography, take a look at [Singh 00], an approachable book by Simon Singh, a science journalist rather than an academic.

REVIEW QUESTIONS 1. Why are layers important in subsystem design? Choose all options that apply. (a) (b) (c) (d)

They make it easier to change the implementation They reduce the number of classes in the implementation They increase reuse They reduce complexity

2. If two customers, on opposite sides of the world, wish to purchase the last ticket available for a concert, what would be a good approach for allocating the ticket? Choose only one option. (a) Introduce an extra business rule that combines a query for ticket availability with a temporary reservation. (b) Make the customers take part in a software ‘race’ to get the ticket. (c) Don’t allow the last ticket to be sold, because that would be unfair to one of the customers.

Review Questions

3. With reference to Figure 6.13, what is Z? Choose only one option. (a) (b) (c) (d) (e)

A class. An event. An interface. A boundary. A property.

Y

X Z

Figure 8.20: For use with Review Question 3

ANSWERS TO REVIEW QUESTIONS 1. Layers are important in subsystem design because: a. They make it easier to change the implementation c. They increase reuse d. They reduce complexity 2. If two customers, on opposite sides of the world, wish to purchase the last ticket available for a concert, a good approach would be to a. introduce an extra business rule that combines a query for ticket availability with a temporary reservation. 3. With reference to Figure 8.20, Z is c. An interface.

239

9

Choosing Technologies

T

his chapter identiﬁes the major technologies available for the client side and the server side and helps to ensure that you know how to make informed choices.

Learning Objectives Understand client-side technologies.

Understand protocols for connecting clients and servers.

Understand server-side technologies.

Understand network technologies.

•

•

Chapter Outline 9.1 Introduction 9.2 Client Tier Technologies 9.3 Client Tier to Middle Tier Protocols 9.4 Middle Tier Technologies 9.5 Middle Tier to Data Tier Technologies 9.6 Other Technologies 9.7 Typical Front-End Conﬁgurations 9.8 Back-End Conﬁgurations 9.9 Java E-Commerce Conﬁguration 9.10 UML Packages 9.11 Summary Further Reading Review Questions Answers to Review Questions

9

242

Chapter 9

9.1 INTRODUCTION Although we have a full set of requirements and analysis documents and even an initial architecture diagram, we still haven’t made any choices about the implementation technologies that we’re going to use. As already mentioned, the longer we delay this decision, the less sensitive our system will be to changes in the future. On the other hand, the longer we wait before we choose technologies, the less chance we will have of exploiting the ﬁner points of those technologies. Making technology decisions at this stage in the development (before detailed design) is a good compromise. Once we have made the decisions, we can still make informed choices as we move forward about how much we want to tie ourselves to the unique features of any particular technology. All the diagrams in this book stick to language-independent UML. For example, the standard UML operation signature has been used, rather than switching to something more speciﬁc, such as Java syntax. In contrast, Java primitives have been shown in the diagrams, rather than UML primitives, because they’re more compact and because it is clearer that we’re using a primitive value rather than an object value.

9.2 CLIENT TIER TECHNOLOGIES Let’s look at the software that runs on clients in a multi-tier system (the main focus of this book). We have two main choices: we can launch either a purpose-built application (program, executable, or whatever you prefer to call it) or a Web browser that hosts the client software. The kind of client applications that we can run include: • • • •

Human to human communications: e-mail, instant messaging, USENET news, chat. File transfer or ﬁle swapping. Remote login. Proprietary applications (anything that doesn’t ﬁt into a more general category, for example, a multiuser ﬂight simulator). Clients hosted in a Web browser can use the following technologies:

• HTML forms. • JavaScript. • Proprietary plug-ins (anything that isn’t one of the other general categories, for example, interactive animation using Flash). • ActiveX controls. • Java applets.

Client Tier Technologies

All of the above technologies use some kind of protocol (for example, IMAP for e-mail, AIM for instant messaging and HTTP/CGI for HTML forms) to communicate with at least one other machine (mail server, messaging server or Web server). Generally speaking, applications and browser plug-ins can be written in whichever language you prefer (Java, C++, Eiffel, Fortran, or COBOL). This was always the case for applications; for browser plug-ins, you need a compiler for your chosen language that’s willing to put the machine code into a DLL (Windows) or shared library (Unix). Client applications have the disadvantage that they require installation of software on the client machine before they can be used. However, for some purposes (such as desktop publishing) they’re a good choice. A Web browser, an application in its own right, is especially useful for client software because it can be enhanced by the multi-tier developer to behave in ways never dreamt of by the developer of the browser itself. For example, a Web browser can be used to read news bulletins, as HTML pages, from a global broadcaster; it can also be used to check bank account details using a Java applet. And all this is done without any prior installation on the client machine (apart from the browser itself of course). For a large client tier written in Java, it may be impractical to load an applet over the network – in such cases, local installation of the applet (or a Java application) may be necessary. Section 8.4.4 suggested that network computers were a good idea, regardless of the size of application loaded over the network. The reality is rather more subtle. Normally, a network computer uses Internet technology (over an intranet) to get hold of its software. Internet-style networks, by their very nature, have bottlenecks that make the frequent loading of large programs impractical. Although this problem can be alleviated by having a good local cache, the best kind of network computer is still a disk-less Unix workstation, which suffers far fewer bottlenecks. For the time being, this means that ‘If you want to load a large program over the Internet, don’t. If you want to load a program over an intranet, make sure it’s not very large (<2 MB)’. All is not lost, however: it turns out that a useful thin-client GUI can be squeezed into a 100KB Java applet, which is practical even for the Internet. Each browser technology has its own advantages and disadvantages. For example: • HTML is visually rich and widely supported, but HTML forms are primitive and they’re not automatically validated on the client. Also, being forced to step through a dozen pages, each of which ﬂashes to white before it’s drawn, is not the nicest way of interacting. • JavaScript permits some client-side programming (for example, data validation in HTML forms). But JavaScript is interpreted (and therefore slower than compiled code), is not pure (in the object-oriented sense) and different browsers provide different levels of support (which makes coding awkward). • In theory, plug-ins can provide any kind of client interaction you wish to implement. However, plug-ins often need to be downloaded and installed the ﬁrst time they’re used; each one requires different programming expertise; and they must be ported by the

243

244

Chapter 9

supplier to each new platform (operating system/CPU combination) so they may not be supported by every client. • ActiveX controls are 32-bit Windows binaries hosted by a Web browser. Although this gives them the same possibilities as plug-ins and Java applets, they can only run on Windows. • Java is a simple, pure, object-oriented language, so potentially it offers the best solution. Java also provides a measure of safety because it prohibits access to resources on the local machine without explicit and detailed authorization by the user. Partly because of corporate chicanery, at the time of writing most Web browsers only support an old version of Java – for a full-featured version of Java, you need the Java PlugIn from Sun (Java masquerading as a browser plug-in) or you need to buy a PC with Java preinstalled. In principle, Java is the best choice for implementing clients in a three-tier system. Because of the lack of out-of-the-box support in Web browsers, most developers opt for HTML forms. (To be fair, the problem of browser support also plagues other client options, such as plug-ins and ActiveX – HTML forms are only practical because they have been around for so long.) Many Web browsers can also host older client applications, including ﬁle transfer, e-mail and USENET news. Most of the important technologies are migrating onto newer devices such as personal digital assistants (PDAs) and mobile phones. TCP/IP is ubiquitous, even on small client devices – it even works over wireless connections. Thus, porting a three-tier system to new kinds of devices will normally just involve redesigning the user interface to make it smaller (and more primitive).

9.3 CLIENT TIER TO MIDDLE TIER PROTOCOLS Client software, whether running as an application or inside a Web browser, has to communicate with a server using some kind of protocol. Most protocols are layered: at the bottom we have a low-level protocol such as TCP/IP and on top of that we build further protocols, specialized for particular tasks. For example, on top of TCP/IP, we can put the Secure Sockets Layer (SSL) which encrypts and decrypts the information for the sake of privacy and integrity. On top of SSL we might run Secure HTTP (HTTPS), a secure protocol that allows a client to request a document by URI and get back the contents of that document. It’s quite reasonable to have multiple layers. For example, Java has a mechanism called Remote Method Invocation (RMI) that allows an object to send a message to an object running on a separate machine – the message is sent using a protocol called Java Remote Method Protocol (JRMP). In order to get through Internet ﬁrewalls, RMI is prepared to

Client Tier to Middle Tier Protocols

piggyback HTTP if necessary. So, when an object sends an RMI message to another object, we may end up with the run-time situation shown in Figure 9.1 – this shows a message being encoded using JRMP, then HTTPS, then SSL, then TCP/IP; on the server side, the message is unwrapped and sent to the intended recipient. The reply is sent back via the reverse path.

aClient

aServant

JRMP HTTPS (Optional) SSL (Optional) TCP/IP

Figure 9.1: How RMI can use a network

Commonly-used protocols fall into two categories: specialized and general. Specialized protocols include: • • • • • •

IMAP (e-mail). AIM (AOL Instant Messaging). NNTP (USENET news). HTTP/CGI (HTML forms). FTP (ﬁle transfer). Telnet (remote login). General protocols (adaptable to many tasks) include:

• TCP/IP (low level transport, also known as sockets). • JRMP (for Java-to-Java communications). • IIOP (for CORBA communications, similar to RMI but multiple implementation languages). Developers usually use a higher level abstraction, courtesy of their run-time system and libraries. For example, RMI and CORBA programmers just send messages to objects – the RMI or CORBA code does all the wrapping and unwrapping invisibly; programmers of HTML forms simply design the layout of their forms – the Common Gateway Interface (CGI) mechanism arranges for the form data to be passed to executable code on the server.

245

246

Chapter 9

9.4 MIDDLE TIER TECHNOLOGIES Having seen the ways of getting to the ﬁrst line of servers, we need to decide what kind of software will take over from there. Server applications are typically multi-threaded pieces of code, designed for high throughput (capable of handling thousands, or millions, of clients simultaneously). A server application listens on some port (connection point) for clients to connect. As far as styles of software are concerned, the server situation is similar to that on the client. On the client, we can run stand-alone applications or code hosted by a Web browser. On the (middle tier) server, we can run stand-alone applications or we can run a Web server and put our code inside that. Stand-alone applications include: • • • • • •

Mail, messaging, news and chat servers. FTP daemon. Telnet daemon. RMI Registry (a look-up mechanism for RMI objects). CORBA naming service (a look-up mechanism for CORBA objects). Java Naming and Directory Interface (JNDI) server (a general name-to-thing mapping service that can be used instead of an RMI registry, a CORBA naming service, a user registry, etc.). • Proprietary server (for example, a process hosting CORBA or RMI objects, an EJB client, a .Net client) Server code that can be hosted by a Web server includes: • Java Server Pages (JSPs), for building Web pages on-the-ﬂy. • Active Server Pages (ASPs), similar to JSPs but coding is typically Visual Basic rather than Java. • CGI scripts (these can be interpreted ﬁles, written in languages such as PERL, or executable programs). • Servlets (Java server objects that can be accessed by Java applets, JSPs or HTML forms). RMI registries and CORBA naming services allow RMI and CORBA clients to ﬁnd their server objects by name (or they can use JNDI). CGI scripts are either textual ﬁles written in some command language, such as PERL, or native executables, compiled from a programming language in the normal way. A CGI script is normally invoked by an HTML form: the Web server strips the data out of the form and passes it to the CGI script as environment variables; the script places its result on standard

Middle Tier to Data Tier Technologies

output, which is returned to the client browser and displayed according to its type (HTML page, image, audio ﬁle, etc.). Servlets are Java objects instantiated on-demand by a Web server. A servlet is usually passed data from an HTML form for processing: as with CGI scripts, the result returned by the servlet speciﬁes what the user sees next in their browser. A servlet is a platform-independent, fast alternative to a CGI script. JSPs are text ﬁles containing raw HTML interspersed with Java code. When ﬁrst invoked a JSP is translated into a servlet: the raw HTML is replaced with print statements and the Java code is included verbatim; the servlet is then compiled and invoked. JSPs are normally used to personalize Web pages – by inserting a bank statement for the customer who’s currently logged in, for example. They can be invoked directly or via servlets. ASPs are similar to JSPs but they use Microsoft technologies, so they’re not as portable.

9.5 MIDDLE TIER TO DATA TIER TECHNOLOGIES So far, we have run some code on the client, launched ourselves onto the middle tier via some protocol and invoked some code on the middle tier. What do we do next? Well, the usual answer, in a three-tier context, is that we access the data tier. There are a number of ways of doing this: • Include database-client code on our middle tier so that we can access a DBMS running on the data tier. With Java, we can do this generically with the help of the Java Database Connectivity (JDBC) mechanism [Campione et al. 98]. • Communicate with the data tier using any of the client-to-middle-tier technologies already discussed. After all, as far as the data tier is concerned, the middle tier is just another client. • Do something proprietary, such as access a server running on a data-tier machine or run some code directly on the middle tier (a two-tier conﬁguration). • Access the data tier using some non-TCP/IP protocol (we will usually only do this in order to access a legacy system). • Include Enterprise Java Beans (EJB) client code in our middle-tier server, then access the data tier via EJBs (Java objects that are typically used to provide data and process services on an intranet). • Include .Net client code in our middle tier server. The .Net framework is Microsoft’s competitor to the EJB framework (and other parts of J2EE). Since this book uses Java-centric examples, no further detail of .Net is given.

247

248

Chapter 9

Figure 9.2 shows how all the technologies discussed so far ﬁt together, in terms of where they normally sit in a three-tier system.

Mail Messaging News Internet Relay Chat File transfer tool Remote login (telnet) Proprietary tool Web browser - HTML/CGI - JavaScript - Plug-in - Applet

IM AP ,A IM .N NT HT P, IRC TP ,T ,F CP TP /IP ,T , JR EL NE MP T, , II OP

OP

DBMS JDBC net server Database client Proprietary data server EJB data source

II et, Cn B JD ol, toc o r p /IP DB CP ry ,T a t P e M pri JR Pro

Mail server Messaging server Internet Relay Chat server File transfer daemon rlogin/telnet daemon Proprietary server (e.g. CORBA, RMI, EJB) JNDI server CORBA naming service Web server - HTML - CGI scripts - Applet code - Servlets

Figure 9.2: Summary of three-tier technologies

9.6 OTHER TECHNOLOGIES There are a number of other technologies that deserve a mention at this point. These technologies have been excluded from the discussion so far, either because they’re relatively new and may not survive or because they simply provide alternatives to what has already been discussed. • Authentication: When users are accessing a system over a network, it is desirable to verify who they are. We can either invent our own mechanism for doing this or we can

Other Technologies

•

•

•

•

use existing techniques – for example, a servlet can instruct a Web browser to display a login screen before proceeding. However, what we would really like is a mechanism that forces the user to log in and then preserves their identity throughout the system: from the Web browser, through the middle tier services and onto the back end. This desirable goal is referred to as single sign-on (or global sign-on if we can cross to other domains). Single sign-on for Java is part of J2EE. For global sign-on, emerging technologies include Microsoft .Net Passport (www.microsoft.com) and those under development by the Liberty Alliance (www.projectliberty.org). XML: With networked systems, we frequently need to transmit business data from one machine to another. The eXtensible Markup Language (XML) [Yergeau et al. 99] is a textual language for describing data as structured name–value pairs. Since XML documents are simply streams of characters, they can be transmitted over whichever protocol we happen to have chosen for our system, HTTP or RMI, for example. For our XML documents, we can invent whatever structure and name–value pairs we like. For XML to be truly useful, when one of our documents arrives at another machine, the target machine has to be aware of the meaning of the data, not just its structure. For example, if we deﬁne an XML document with a value called NETAMOUNT, our document will be of little use unless the recipient knows that the document is an invoice and that NETAMOUNT is the amount to pay before sales tax. For this reason, standards work is underway to deﬁne XML document types for common business information. In the absence of standard document deﬁnitions, XML is still useful because it allows us to use the same libraries for parsing and generating all our documents (libraries are available for Java, for example). XML can be used for storing textual information, as well as transmitting it; for example, we can store spreadsheet documents in XML format. Events and messages: Events are a common mechanism for broadcasting information to interested parties. The trouble with events is that they don’t work well over a network, because it can be tricky to make them thread-safe and because they’re inefﬁcient (from the server point of view). A better way to broadcast over a network is to use machineto-machine messaging. Messages are nuggets of information that can be broadcast to interested clients: once the server has initiated the broadcast, it plays no further part in the delivery of the information; failure to deliver to a particular client has no adverse effect (messages can also be sent in guaranteed-delivery mode). For more information on messaging, investigate the Java Messaging Service (JMS) [Bodoff et al. 02]. SOAP: The Simple Object Access Protocol (SOAP) [W3C 03] is similar to RMI and CORBA. What sets it apart, however, is its use of a protocol based on XML, which makes it a good candidate for becoming the de facto standard of the future. Web services: The idea with web services is that customers (at home or in business) will pay to have their information stored and processed on fat servers, accessible over the Internet or extranets. A simple example of a Web service is browser-based e-mail;

249

250

Chapter 9

a more complex example is the processing, managing and printing of digital photos. Web services are deployed as Internet-based three-tier systems, using TCP/IP and XMLrelated technologies, such as SOAP. For more information check out the Web Services Interoperability Organization (www.ws-i.org) and the Java-related efforts led by Sun (java.sun.com).

9.7 TYPICAL FRONT-END CONFIGURATIONS Having seen the many technologies available for tiers in a multi-tier system, let’s take a look now at some typical conﬁgurations for communication between the client tier and the middle tier. While reading the next few topics, bear in mind that getting from the client tier to the middle tier is a user interface exercise: by keeping all the complicated business logic on the middle tier, we can happily choose any or all of these front-end conﬁgurations, according to personal preference or the time available. In Chapter 10, we’ll see how to design the middle and data tiers so that we can swap our front ends in and out as we wish.

9.7.1 HTML/CGI-with-Scripts Originally, HTML was a simple page markup language for multimedia documents with embedded hyperlinks [W3C 99]. A Web browser session would be started by a user typing in a destination URI, such as http://www.blueskyuniversity.com/salespitch.html, and hitting the return key (or by selecting a bookmark or favorite). The browser would then parse the URI, detect that HTTP is being used to connect to the Web server and pull out the Web server’s IP address (www.blueskyuniversity.com). The browser would then contact the Web server and pass it an HTTP request to fetch the ﬁle salespitch.html. The Web server would retrieve the corresponding ﬁle, a Web page in this case, and return it to the browser for display. This basic HTTP mechanism was ﬁne for browsing distributed multimedia documents, but it provided no way for the user to enter data, such as which book they would like to purchase. The Common Gateway Interface (CGI) addresses this issue. With CGI, the HTML displayed in a browser contains one or more forms (text ﬁelds, drop-down lists and buttons) for the user to ﬁll in. All the user has to do is enter data into the form and click a button (typically labeled Submit or Proceed to Checkout). When the button is pressed, the Web browser extracts name–value pairs from the form and passes them to the Web server with the name of the CGI script (command ﬁle, interpreted ﬁle or executable) that should process the data. All the information about the script and the name–value pairs can be parceled up by the Web browser as a URI: http://www.blueskyuniversity.com/cgi-bin/buy.pl?b=Gemma&q=2

Typical Front-End Conﬁgurations

(The name–value pairs can also travel across in the body of the HTTP request, but the end result is the same.) When the Web server receives the request, it detects that it’s being asked to run a process because of the /cgi-bin part of the URI – this speciﬁes a location on the server where the scripts are kept. The Web server ﬁnds the script to run (a PERL script called buy.pl in this case), starts it up and then passes it the name–value pairs as environment variables (operating system values that can be retrieved by a process). In the book example, above, we have two name–value pairs representing a book and a quantity, respectively, b=Gemma and q=2. The script can now do any processing it needs to do, based on the values it was passed (presumably, in this case, arranging for the delivery of two books and the debiting of the customer’s credit card, using details already entered). Once the script has ﬁnished, anything that it writes to standard output is passed back to the client browser as the result of the interaction – for example, it could be a new HTML page, a picture or a sound clip. In the book example, we would probably send back an order conﬁrmation page. The HTML/CGI conﬁguration is summarized in Figure 9.3. This shows a Web server providing access to HTML ﬁles, scripts and other media ﬁles, while the client hosts a Web browser that displays HTML pages and provides data capture via CGI forms.

Web browser - HTML - CGI forms

HTTP

Web server - HTML - CGI scripts - Media files

Figure 9.3: Conﬁguration using HTML/CGI-with-scripts

Thus, CGI allows us to collect form data from the client for processing on the middle tier. The main advantage of CGI is that it’s supported everywhere, regardless of how old the client’s Web browser is. Thus, any machine can be a client in a three-tier system simply by having a Web browser installed.

251

252

Chapter 9

CGI has a number of disadvantages: • Poor interactivity: CGI forms have a limited number of input components: text entry ﬁelds, lists, drop down lists and buttons. The sophisticated modern user will be used to notebooks, toolbars, spin buttons, menus, and so forth. Using a client-side scripting language such as JavaScript can alleviate some of these problems, at great cost in time and effort to the developer. But, even then, the end result is not perfect. Worse still, each click of a Submit button might take the user to another page, which causes the whole of the browser window to be cleared to the background color and repainted. In contrast, a proper GUI paints much more quickly. • Slow speed: The user has to wait for the contents of the form to be dispatched to the server, for the process to be started and run to completion, and for the results to be brought back and painted. Thus, the user gets the impression of slow interaction. • No data validation on the client: We can use JavaScript on the client to improve things a little by validating the form before dispatch. However, we still have to do checks in the server to protect ourselves from hackers, so all we can hope to do with client-side validation is to increase the number of correct forms that hit the server. • Overloaded server: On the server, we have a serious problem because we’re trying to start one process per form. This problem is so serious that we simply couldn’t contemplate processing thousands of forms simultaneously. We can use a tool called FAST-CGI that allows us to keep a process in memory to avoid starting one for each request but we still need a separate copy of the program’s data area for each request. • Scripts may not be portable: Historically, programming languages, scripting languages and command languages have not been entirely portable. Therefore, if we decide that we want to change our middle tier servers from, say, Windows to Unix, there will be some porting effort. • No security: By default, name–value pairs are passed to the Web server unencrypted, so personal information can be read in transit. To get round this, we can run HTTP over SSL (look for https:// at the start of the URI before you submit your form). But this is extra conﬁguration that we must do ourselves (and perhaps pay money for).

9.7.2 HTML/CGI-with-Servlets Courtesy of some clever design by Sun and partners, the HTML/CGI-with-servlets conﬁguration is nearly identical to the HTML/CGI-with-scripts conﬁguration (see Figure 9.4). The client has the same Web browser, HTML pages and CGI forms. The difference is in the URI: rather than nominating a script to run, we specify a servlet (a Java object) to instantiate: http://www.blueskyuniversity.com/servlet/BuyServlet?b=Gemma&q=2

Here, the Web server spots that we want to run a servlet because of the /servlet part of the URI. Next comes the class name of the servlet (BuyServlet). The ﬁrst time the Web server

Typical Front-End Conﬁgurations

Web browser - HTML - CGI forms

HTTP

Web server - HTML - Servlets - Media files

Figure 9.4: Conﬁguration using HTML/CGI-with-servlets

receives a request for a particular servlet, the servlet is instantiated and told to process the request. Thereafter, the servlet stays alive and all subsequent requests go through the same servlet. (Each request is run through the servlet in a separate thread of execution. So, rather than running scripts as separate processes, we have the Web server passing multiple client requests through a single Java object.) To pass a request to a servlet, the Web server uses one of a handful of standard messages, doPost(:HttpServletRequest,:HttpServletResponse) being the most common; the name–value pairs are inside HttpServletRequest. When the doPost method has ﬁnished its back-end processing, the response is put into HttpServletResponse. The Web server extracts the contents of HttpServletResponse and dispatches it back to the Web browser in the normal way. Servlets have the following advantages over scripts: • Performance: A servlet handles multiple requests simultaneously without multiple processes or multiple data areas, so requests are serviced more efﬁciently. Also, servlets normally run faster than interpreted programs written in languages such as PERL. • Scalability: Since multiple servlets can be handled by a single (Web server) process, we can manage millions of simultaneous client requests without overloading the server. • Portability: If we change our operating system, the only porting effort is to install the appropriate version of Java on the new platform. The servlets themselves remain unchanged. • Ease of use: Since servlets are standard Java objects, servlet developers have access to a huge library of reusable objects for common computing tasks, so there’s less coding to do than with scripts.

253

254

Chapter 9

9.7.3 RMI Remote Method Invocation (RMI) [Campione et al. 98] is a fancy term that means ‘Java objects sending messages to other Java objects over a network’. RMI is part of the standard Java libraries, so it’s available to anyone who installs Java 2 Standard Edition (J2SE) or Enterprise Edition (J2EE). The RMI designers have worked hard to make RMI simple to use and remote objects are almost as easy to use as local objects. The only differences are that: • Server objects must be deployed on the machines that are to be contacted over RMI; they may provide all the services needed by remote clients or they may simply be factories for other objects. • Any machine hosting server objects must run a naming service, a server process running on a well-known port that allows remote machines to look up server objects by name. RMI has its own naming service called a registry; however, it is better to employ the pluggable Java Naming and Directory Interface (JNDI), so that you can exploit whatever naming service is available. • Any message sent over the network may throw an exception, RemoteException: this ensures that client programmers can’t ignore the fact that they’re stepping outside of their local Java process. Apart from these minor restrictions, the developer can use RMI to set up any client–server or distributed architecture. Each host in an RMI conﬁguration can be implemented as an applet or an application. The applet scenario has the advantage that, as long as we deploy a Web server somewhere on the network, the client doesn’t need any of our system software installed beforehand – all the client software can be loaded as it is needed via the Web browser and Web server. For applications, we have to do one of the following: • Deploy the client software to each client machine. Whenever we update the client software, we need to redeploy on all clients. This option is best left for situations where the system software for the client is large (what large means depends on the speed of the network you’re using). • Deploy a Web server and then use a small piece of bootstrap software to load the client code. The bootstrap software still needs to be installed on each client, but at least we guarantee that the client gets the latest version of the software on each startup. • Use Java WebStart to load the client software as needed. Java WebStart is a tool provided by Sun to enable the dynamic loading of Java applications (a mechanism normally reserved for applets). To use Java WebStart, we need to deploy a Web server and then make sure that each client has a Web browser with Java WebStart and the Java PlugIn (part of J2SE)

Typical Front-End Conﬁgurations

installed. This is the simplest option if we want to run clients as applications rather than as applets. Figure 9.5 shows the front-end conﬁguration for an RMI applet client. On the middle tier, we must run some kind of server process that contains our system-speciﬁc code in addition to the generic code that runs as part of the naming service.

Web browser - HTML - Applets

JRMP

Web server - HTML - Applet code - Media files RMI registry / JNDI server Proprietary RMI application

Figure 9.5: RMI used from applets

9.7.4 CORBA The Common Object Request Broker Architecture (CORBA) [OMG 04] was developed by the Object Management Group, the industry consortium dedicated to the development of open object standards that also owns UML. CORBA is similar to RMI but it predates RMI, so it has a stronger foothold in industry. The main differences between CORBA and RMI are: • CORBA is multi-language: host software can be written in C++, Eiffel, C#, Java and even non-object-oriented languages such as C or COBOL (the server software will still appear object-oriented to clients). • CORBA has its own dedicated naming service (although this can be used via JNDI). • Java hosts receive unchecked exceptions from CORBA but checked exceptions from developer code – this can cause confusion. (A checked exception must be handled by the programmer, whereas an unchecked exception need not be.) • CORBA costs money (the free implementation that comes with Java is suitable for development and testing purposes). • RMI has access to the entire Java 2 platform (Standard Edition, at least). This is useful for gaining access to media ﬁles, for example.

255

256

Chapter 9

CORBA, because of its generality, is more complicated than RMI. RMI, on the other hand, is more powerful (it supports the dynamic exchange of object graphs and compiled classes). Also, RMI performance can be optimized for each new application. Other than that, the Java programmer should regard CORBA and RMI as alternatives. Non-Java programmers should regard CORBA as the best choice for simple message-sending over the network (it is preferable to proprietary, non-portable alternatives). Sun has developed a bridging protocol called RMI-over-IIOP that allows RMI hosts to be mixed with CORBA hosts, with a few restrictions. Figure 9.6 shows the front-end conﬁguration for applets accessing CORBA. For non-Java hosts, CORBA client software must be deployed on each client (and redeployed for each new release). CORBA hosts written in Java have the same deployment options as RMI hosts.

Web server - HTML - Applet code Web browser - HTML - Applets

IIOP CORBA naming service Proprietary CORBA application

Figure 9.6: CORBA used from applets

9.7.5 EJB Enterprise Java Beans (EJB), part of J2EE [Bodoff et al. 02], is a framework for distributed Java systems with full transaction management, security and persistence support. In Java terms, a bean is an object with a well-known style of interface. EJB implementations are available from many industry leaders such as IBM, Sun and BEA Systems. There is also a reference implementation available from Sun, so that you can develop and test EJB software without buying an implementation. The EJB framework is, in many ways, a competitor to Microsoft’s .Net strategy (especially if we include all the other facilities of J2EE). .Net, of course, is proprietary and not very portable, whereas EJB is open and portable. Detailed discussion of the EJB framework is beyond the scope of this book. Sufﬁce to say that EJBs come in three main varieties:

Typical Front-End Conﬁgurations

• Entity beans, which correspond to entity objects in the Jacobson/UML sense: business objects that have business information and business behavior. Entity beans can be stored automatically by an EJB implementation, using a relational database of your choice, or the developer can elect to store the beans using proprietary code (for integration with legacy systems and object-oriented databases). Access to entity bean data is subject to transactions: the developer can settle for the default or choose from a variety of options (representing different speed/accuracy trade-offs). For the sake of portability, all transaction semantics are speciﬁed by the framework while the implementor of each EJB run-time system has to make sure that the semantics are supported by the implementation. For efﬁciency reasons, messages are sent to entity beans locally, rather than over the network. • Session beans, which manage business tasks on behalf of EJB clients. A client sends a message over the network to a session bean, which satisﬁes the request using entity beans and other session beans. A session bean provides default transaction behavior for all the entity beans and session beans that it accesses. Alternatively, the client can control when transactions start and ﬁnish (and, to a limited extent, how they should be propagated to other beans). • Message-driven beans, which integrate seamlessly with industrial strength machine-tomachine messaging implementations. Messages can be sent peer-to-peer or broadcast to several machines at once, under full transaction control. Figure 9.7 illustrates the conﬁguration for an EJB application front end. In order for EJBs to work, they must be dropped into an Enterprise Java Server (EJS), often referred to using the more general term application server. This is different from the RMI and CORBA cases,

Web browser - HTML - Java applications (EJB clients)

IIOP/JRMP

Enterprise Java Server - Proprietary EJBs JNDI server

Figure 9.7: An EJB application front end

257

258

Chapter 9

where we had to provide our own server to contain our business objects. As with RMI and Java-based CORBA, various alternative conﬁgurations are possible.

9.8 BACK-END CONFIGURATIONS So far, we’ve seen how clients of a three-tier system can be implemented using CGI, RMI, CORBA or EJBs. For Internet applications, the most common choice is HTML/CGI. For intranets, we can choose our preferred mechanism, although, for the sake of maintenance, we should strive to keep the middle tier in a three-tier system applicable to all types of client, regardless of which type of access is implemented ﬁrst. (If we have a controlled user base – users who are prepared to install the correct level of Java – we can choose whichever front-end technology we prefer, even for Internet use.) Any of the conﬁgurations described can be used as the front end for a two-tier system. So, what happens when we arrive at the middle tier? The simplest solution is to install a DBMS client on the middle tier and access it from whatever server software we have chosen to deploy (scripts, servlets, RMI server, CORBA server or Enterprise Java Server). Access to the database client may be possible within the same process as our middle-tier code, or it may have to be in a separate process. Alternatively, we can arrange for our server software to act as a .Net client or an EJB client. Since the Enterprise Java Server or .Net server that we access from the middle tier provides business services, logically it belongs in the middle tier; however, all of these technologies permit deployment across multiple machines in many different ways so it doesn’t matter where the server actually resides.

9.9 JAVA E-COMMERCE CONFIGURATION Because e-commerce systems need to attract the most casual of Internet users, we’re really stuck with the HTML/CGI front end as the default mechanism. If we deploy any other kind of front end, we risk losing customers: most customers, when confronted with ‘You must install such and such to use this site’ will simply go elsewhere. Front ends not hosted by a Web browser are also out of the question: it is not reasonable to expect e-commerce customers to install software for our beneﬁt. With the help of JSPs (or ASPs, for the Microsoft-minded), we can produce an interaction that is tailored to each user (although we can’t solve the repainting and slow-round-trip problems inherent in browsers). Once we’ve reached the middle tier, we can choose any technology we like to access the data tier and other parts of the middle tier. So, how do we combine CGI, JSPs, servlets, EJBs and DBMSs into a coherent, scalable whole? The basic idea is that we use CGI to access servlets that provide the business services;

Java E-Commerce Conﬁguration

the servlets rely on reusable EJBs to do most of their work; the EJBs rely on a DBMS to store enterprise data; ﬁnally, the servlets pass the result data to JSPs to build the personalized Web pages that are sent back to the client. The full picture is illustrated in Figure 9.8. (Nonstandard, UML-style keywords have been used to show how each object ﬁts into the EJB framework.)

Enter your details below and click the button to log on:

invokes

User Name wilma37 Password

***********

<<servlet>> :Authentication Servlet

Log On

Home About Contact Help

<<sessionBean>> :Authentication Session

Hello, Wilma. What would you like to do next?

Play a game Chat to a friend Log out

<<entityBean>> :Customer paints

<<jsp>> :Logged On

Home About Contact Help

Client Tier

Middle Tier

Data Tier

Figure 9.8: Java e-commerce technologies in action

In the scenario illustrated, Wilma navigates to the logon page of our site, so that she can access services that are only available to registered customers. Wilma ﬁlls in her user name – Wilma37 – and password and clicks the Log On button. Wilma’s form data is passed to the AuthenticationServlet for processing. The AuthenticationServlet invokes an EJB session bean – the AuthenticationSession – to check the user name and password. The AuthenticationSession bean asks the EJB run-time system to ﬁnd the CustomerEntity bean that has the user name Wilma37 (if there is no such customer, an error is signaled to the

259

260

Chapter 9

servlet). The AuthenticationSession then compares the CustomerEntity’s password with the one passed in by the servlet. If the passwords match, the AuthenticationSession returns a session token (a unique identifying value) to the AuthenticationServlet; if they don’t match, the AuthenticationSession signals an error to the servlet. In the case of a successful logon, the AuthenticationServlet associates the session token with the browser session (usually as a cookie, a small piece of information that gets passed back to the browser), along with any other relevant information (such as the key of the CustomerEntity). Next, the servlet forwards the request to a JSP, passing it a Java object

*

*

<<device>> CootHTMLClient <<device>> CootGUI Client

<<ExecutionEnvironment>> WebBrowser

JRMP

HTTP

2

<<device>> CootServer

<<ExecutionEnvironment>> WebServer

<<ExecutionEnvironment>> CootBusinessServer

2

<<device>> DBServer

<<ExecutionEnvironment>> DBMS

Figure 9.9: A more detailed deployment diagram for iCoot

Java E-Commerce Conﬁguration

containing result data (such as the real name of Wilma37). The JSP runs its embedded Java code to interleave personalized data with the static HTML. Once the completed page has been sent back to the browser and displayed, the business transaction (use case) is complete. In the event of an unsuccessful logon, the error raised by the AuthenticationSession is detected by the AuthenticationServlet. The servlet forwards this request to an error page, which might be a static HTML page containing a generic message, such as User name or password incorrect, or a dynamic page (built by a JSP) with something more speciﬁc, such as The password you entered was incorrect. Returning an error page to the browser marks an alternative end to the business transaction (use case).

Case Study iCoot conﬁguration Because of its simplicity, portability and light loading of the client, the HTML/CGIwith-servlets conﬁguration will be used for the ﬁrst increment of iCoot. A later increment will also provide a proper Java client so that customers can use a GUI for elegant and fast access to the system via RMI. On the server side, a proprietary business server that accesses a relational database using JDBC will be used. The business server will use EJB session beans to provide access to Java transactions but it won’t use EJB entity beans because the developers have not yet been trained in their use. We can add these architectural decisions to the topological deployment diagram that we saw earlier (in Section 8.4.9). Figure 9.9 shows processes as subnodes, each marked with the UML keyword <<ExecutionEnvironment>>. The communication paths now show navigability – the path name can be used to indicate the communication protocol. The deployment survey for this extended picture is: The iCoot data tier comprises two database servers (which we have called Having two such nodes improves throughput and reliability. Each DBServer hosts a DBMS process for managing access to data. The middle tier, which communicates with the data tier, consists of two server machines (CootServer), again duplicated for the sake of reliability and throughput. Each CootServer hosts a CootBusinessServer (for handling business requests) and a WebServer (for handling static HTML content and forwarding business requests to the CootBusinessServer). Data access for the CootBusinessServer is provided by the DBMS. Because they’re proprietary to the products that we select, the communication protocols between the WebServer and the CootBusinessServer and between the CootBusinessServer and the DBMS are not speciﬁed. DBServer).

261

262

Chapter 9

Case Study (cont’d) Each CootServer can be accessed simultaneously by any number of CootHTMLClient nodes. Each CootHTMLClient hosts a WebBrowser, which accesses one of the WebServer nodes using HTTP. Eventually, we will also provide access from CootGUIClient nodes. Each CootGUIClient will access one of the CootServer nodes, using JRMP. Because the mechanism that allows such requests to get into the CootBusinessServer is the subject of a future increment, no details are given. Nor is any detail given for the CootGUIClient processes.

9.10 UML PACKAGES The UML concept of a package allows us to group related classes. The package diagram in Figure 9.10 shows each package as a box with a tab at the top left corner. The package name appears in bold, either in the middle of the box or, if we want to show the contents of the package, inside the tab. The contents of a package can be classes or other packages. Figure 9.10 also shows a dependency (the dashed open-ended arrow) from one package to another: the implication is that the source package uses something inside the target package. myPackageToo Class1

myPackage

Class2

myNestedPackage

Figure 9.10: UML package diagram

A package can be used to represent: • • • •

a layer a subsystem a reusable library a framework

UML Packages

• classes that should be deployed together • ... From a programming point of view, packages map conveniently to existing language constructs such as Java packages and C++ namespaces. But bear in mind that packages are simply a compile-time concept: they help us to organize code for the beneﬁt of development, deployment and maintenance.

Case Study iCoot packages Figure 9.11 shows the package design for iCoot. In this case, each layer in the system has been mapped to a distinct package. Dependencies on packages in the standard Java library such as java.sql, which is used by the business package for access to the database, have not been shown. If all dependencies had been included, the diagram would have been much larger and much more cluttered. The main iCoot package is labeled com::nowhere. This is shorthand for ‘package nowhere is nested inside package com’. UML borrows this notation for the namespace operator from C++ (when writing Java source code, the ‘::’ is replaced with ‘.’). Despite the use of events in iCoot, there is no dependency from the control package to the packages above. This is because the event listener and event types reside in the control package. The protocol package contains deﬁnitions of the lightweight copy objects, exported by the server layer for simplifying network access to the BusinessLayer. (This is explained further in Chapter 10.)

Many developers will use a package diagram to show layers. However, there are a couple of problems with this approach. Firstly, layers are chosen before any decision about how to organize source code into packages has been made (and the organization may well be different). Secondly, a package diagram doesn’t allow us to show some important information; for example, the existence of the HTTPCGILayer in iCoot is important, but it doesn’t map to any package that we might implement or borrow from a library (it’s simply an indication that we use the HTT P protocol with the CGI extras). Thus, for iCoot, a special-purpose layer diagram, with an accompanying document describing the layer interaction policy, is more appropriate (see Figure 8.14). Package diagrams are also not good at showing horizontal partitions (subsystems); a deployment diagram should be used instead. Figure 9.12 shows the iCoot deployment diagram after system design has been completed, detailing the content of the CootServer processes in terms of UML artifacts and packages.

263

264

Chapter 9

com::nowhere

swing

micro

control

protocol

rmi

servlets

server

business

Figure 9.11: Packages for iCoot

An artifact, in UML, is something that can be deployed, usually a ﬁle. (A ﬁle can be almost anything: for example, a program, a DLL, a folder, some XML data or a README document.) Artifacts are indicated by either a sheet-of-paper icon or the <<artifact>> keyword. The artifacts in Figure 9.12 are iCoot, a folder of static HTML pages; icoot.ear, a compressed archive of servlets, JSPs and EJBs (the Java naming convention .ear is short for ‘enterprise archive’); and cootschema.ddl, a proprietary script for creating the database.

UML Packages

<<device>> CootHTMLClient

*

* <<device>> CootGUI Client

<<ExecutionEnvironment>> WebBrowser

HTTP

JRMP 2

<<device>> CootServer

<<ExecutionEnvironment>> WebServer

<<ExecutionEnvironment>> CootBusinessServer

<<use>> icoot.ear

iCoot

<<manifest>> com::nowhere servlets protocol server business

2

<<device>> DBServer

<<ExecutionEnvironment>> DBMS

cootschema.ddl

Figure 9.12: The iCoot deployment diagram after system design

265

266

Chapter 9

We can show which elements from our model give rise to a particular artifact by adding a dependency from the artifact to the group of elements, labeled <<manifest>> (the artifact is a manifestation of the elements). Here, icoot.ear has been shown as a manifestation of four of the com::nowhere packages. We can show other dependencies on a deployment diagram. For example, with the help of the <<use>> keyword, it’s been shown that iCoot uses icoot.ear.

Case Study Updated iCoot deployment survey The iCoot data tier comprises two database servers (which we have called DBServer). Having two such nodes improves throughput and reliability. Each DBServer hosts a DBMS process for managing access to data. The cootschema.ddl artifact contains commands for creating database tables, in a format speciﬁc to the database being used. This is deployed to each DBMS process, using database-speciﬁc tools (no detail given here). Note that cootschema.ddl contains the schema for the full Coot system, since iCoot and Coot use the same data. The middle tier, which communicates with the data tier, consists of two server machines (CootServer), again duplicated for the sake of reliability and throughput. Each CootServer hosts a CootBusinessServer (for handling business requests) and a WebServer (for handling static HTML content and forwarding business requests to the CootBusinessServer). Data access for the CootBusinessServer is provided by the DBMS. Because they’re proprietary to the products that we select, the communication protocols between the WebServer and the CootBusinessServer and between the CootBusinessServer and the DBMS are not speciﬁed. Within each CootServer, the iCoot folder, containing static HTML pages, is deployed to the WebServer, while the icoot.ear archive is deployed to the CootBusinessServer. The icoot.ear archive contains servlets, JSPs, business objects and (eventually) RMI decorators, from the com::nowhere package. Each CootServer can be accessed simultaneously by any number of CootHTMLClient nodes. Each CootHTMLClient hosts a WebBrowser, which accesses one of the WebServer nodes using HTTP. No artifacts need to be deployed to the CootHTMLClient nodes. Eventually, we will also provide access from CootGUIClient nodes. Each CootGUIClient will access one of the CootServer nodes, using JRMP. Because the mechanism that allows such requests to get into the CootBusinessServer is the subject of a future increment, no details are given. Nor is any detail given for the CootGUIClient processes. The artifacts deployed to the CootGUIClient nodes, if any, are not speciﬁed.

Review Questions

9.11 SUMMARY In this chapter, we’ve looked at: • The major technologies available on the client and server. • The middleware protocols that can be used to connect clients and servers. • Popular technology choices for the front ends to networked systems. • How UML packages can be used to show clusters of related classes on a deployment diagram.

FURTHER READING For an HTML/CGI client tier, accessing servlets on the middle tier, the Struts framework [Robinson and Finkelstein 04] from the open source Apache Software Foundation is well respected. Struts takes most of the drudgery out of implementing an interactive HTML front end. To download Struts and view the documentation, go to www.apache.org. If you’re interested in serious CORBA programming, a comprehensive book that covers both Java and C++ is [Bolton 01]. As usual, the last word on anything technical is the CORBA Speciﬁcation [OMG 04]. For an overview of key multi-tier Java technologies, see [Campione et al. 98], which covers RMI, JDBC, servlets and Java IDL (the Java interface to CORBA). The contents of this book are also available on-line at java.sun.com. In order to learn more about J2EE, including JSPs, servlets, EJBs, transactions, security, XML, XSL, SOAP, JMS and Java ServerFaces (for interactive HTML interfaces) see [Bodoff et al. 02] and the on-line version at java.sun.com. Sun’s web site also has details of the latest Java angle on Web Services.

REVIEW QUESTIONS 1. What are some of the advantages of Java as an implementation technology? Choose all options that apply. (a) (b) (c) (d) (e) (f)

Object-oriented purity. Invented by Microsoft. Network readiness. Portability. Scalability. Security.

267

268

Chapter 9

2. Servlets are a direct replacement for what? Choose only one option. (a) (b) (c) (d) (e)

XML. Applets. ActiveX controls. CGI scripts. CORBA.

ANSWERS TO REVIEW QUESTIONS 1. Some of the advantages of Java as an implementation technology are: (a) (c) (d) (e) (f)

Object-oriented purity. Network readiness. Portability. Scalability. Security.

2. Servlets are a direct replacement for d. CGI scripts.

10 Designing the Subsystems

I

n this chapter, we’ll be looking at subsystem design (also known as detailed design). Having chosen our network topology, our implementation technologies, our subsystems and our layers, it’s time to decide exactly what is going to go inside each subsystem and layer.

Learning Objectives Understand how to design the business layer.

Understand how to design user interfaces.

Understand how to map run-time objects to storable data.

Understand multi-threading.

•

•

Chapter Outline 10.1 Introduction 10.2 Mapping the Analysis Class Model into the Design Class Model 10.3 Handling Persistence with a Relational Database 10.4 Finalizing the User Interfaces 10.5 Designing the Business Services 10.6 Using Patterns, Frameworks and Libraries 10.7 Transactions 10.8 Handling Multiple Activities 10.9 Summary Further Reading Review Questions Answers to Review Questions

10

272

Chapter 10

10.1 INTRODUCTION Because of the size of the subsystem design task, and the creativity that’s unique to every project, we couldn’t hope to write down a step-by-step guide to designing a professional, object-oriented, multi-tier system. Instead, this chapter contains an overview of each of the major tasks. In practice, if you place these tasks at the center of your design effort, everything else will fall into place. First of all, we’ll consider how to design the business layer. This normally involves deciding what objects will populate the layer, how they will be connected and what their interfaces will be. We must transform the business-oriented class model that we developed during analysis into implementation-oriented classes for our chosen programming language. For the purposes of this book, the subsystem design discussion is more or less independent of the programming languages and other technologies that are chosen. However, for the sake of illustration and convenience, some bias is shown towards Java and relational databases. (Despite the Java bias, the design presented is portable to any language that includes pure object-oriented facilities; similarly, the relational database bias won’t tie us to any particular DBMS.) Subsystem design involves transforming a conceptual analysis model into implementable classes, following the strategy laid out in the system design model. In keeping with the system design principles already discussed, subsystem design can proceed as follows: 1. Design the classes and ﬁelds of the business layer, using the analysis class model as a guide. The business layer consists of the entities from our problem domain and the various supporting classes that they need. 2. Decide how any persistent data will be stored and design the storage layout. Persistent data is data that must not vanish when the system is shut down. 3. Finalize the look and feel of the user interface, with reference to the sketches that were produced during the analysis phase. 4. Walk through the system use cases, with reference to the user interface design, noting the business services that must be supported by the middle tier. Business services are questions and commands that a client can send to the server, such as ‘buy a book’ or ‘reserve a car model’. 5. Develop the business services into server objects, whose messages are available over the network. Server objects implement the business services using the business layer, in such a way that different kinds of client can be accommodated. 6. Finalize the measures that are needed to ensure concurrency control and thread safety. concurrency control means using business rules to control access to the system: user names and passwords, reserving tickets before purchase, and so on. Thread safety means making sure that data within a process is not corrupted and that parallel activities do not get in each other’s way.

Mapping the Analysis Class Model into the Design Class Model

10.2 MAPPING THE ANALYSIS CLASS MODEL INTO THE DESIGN CLASS MODEL In moving from analysis to design, some classes will be discarded (controllers, for example) and others will be introduced (such as collection classes that implement multiplicities). For the purposes of this book, designers have freedom to decide how the business objects, boundaries and homes developed up to this point should be turned into implementable code. (This is different from the RUP approach where the analysis class model is massaged into an implementable design model.) For each design class that we come up with, we need to choose the names and types of its ﬁelds. Often, ﬁelds are derived from the attributes or associations discovered during analysis. There is no way we can tell, just from looking at source code, whether a particular ﬁeld started out as an attribute, a composition, an aggregation or an association. This is one reason why the high-level artifacts produced during requirements capture and analysis are important. As well as attributes and associations, we need to consider inheritance. Inheritance relationships don’t need to be mapped into anything new, we just need to decide whether or not to keep them. Because of the complexities involved, inheritance should be handled with care: it is quite reasonable to produce a system with little or no inheritance among its classes. (Any inheritance that we do use is more likely to be introduced during design rather than analysis.)

10.2.1 Mapping Operations But what about the operations? Up to this point, operations have been introduced merely as a way of recording use case realizations – in other words, as the system use cases were simulated, operations were discovered as a side-effect of verifying that the analysis classes would support an implementation. These analysis operations should be ignored for the purposes of design. So, where do our design operations come from? Now that we’re switching to design, we can stick to the programming terms ‘message’ and ‘method’, rather than the UML term ‘operation’. So, where do the messages come from? For most of our objects, regardless of the layer in which they reside, messages will be added for one or more of the following reasons: • To allow client objects to read or change the values of ﬁelds. • To allow client objects to access derived data (for example, as well as a message to read the radius of a Circle, we would expect to be able to read the diameter). • Because our experience or intuition tell us that a particular message might be useful.

273

274

Chapter 10

• Because some framework or pattern that we have decided to use requires certain messages to be present. In addition, when we design the business services for our middle tier, we invent messages for the server objects. As we work out how the business services can be satisﬁed using the business objects, we’re likely to come up with even more messages. In short, as we map the analysis classes, attributes and relationships, into their design counterparts, messages will start to appear from all directions.

10.2.2 Variable Types When we introduce a ﬁeld, we need to decide what type it is: a primitive or a class. For most purposes, we can restrict ourselves to the following types: • The primitive types and simple classes that we would expect to ﬁnd in every objectoriented programming language (for example, int, ﬂoat, boolean, String, List, Set and array, or []). • Classes that we ourselves are designing. • Classes from the patterns and frameworks that we have chosen to use. In some languages, arrays and collections do not mix well – arrays tend to be ﬁxed-size and may not be objects at all. Therefore, base your choice on elegance (collections) or a slight improvement in performance (arrays).

10.2.3 Visibility of Fields As well as providing the names and types of ﬁelds, we must declare their visibility. The visibility of a ﬁeld speciﬁes which pieces of code are allowed to read or modify the value. The following visibilities are enough for most purposes: • Private (shown by - in UML): Only visible within the deﬁning class. • Package (shown by ˜ in UML): Visible within the deﬁning class and to all classes in the same package. • Protected (shown by # in UML): Visible within the deﬁning class, to all classes in the same package, and to all subclasses of the deﬁning class (whether inside or outside the package). • Public (shown by + in UML): Visible everywhere. Normally, if the language permits, developers will make ﬁelds private: apart from the encapsulation beneﬁts, this gives the compiler more optimization opportunities. Sometimes,

Mapping the Analysis Class Model into the Design Class Model

developers will make ﬁelds protected instead, so that developers of subclasses have more opportunities for modifying the behavior of the superclass (although this does increase the coupling between the subclass and the superclass). Fields with package visibility are a bad idea, because they can be accessed by any piece of code in the package without the knowledge of the owning object. Occasionally, for pragmatic reasons (performance and brevity), a developer will give a ﬁeld package access, but only if the chosen language provides some way of making the value read-only (for example, using the keyword ﬁnal in Java). This argument applies even more strongly to public ﬁelds, since they’re also visible outside the package. Visibilities can be applied to messages too. In this case, each message will be: • public, if it’s part of the interface of the package; • package, if it’s implementation code to be used by the class itself and by classes in the same package; • protected, if it’s implementation code to be used by the class itself, by its subclasses and by classes in the same package; • private, if it’s implementation code for use by this class only (which decreases coupling and allows the compiler to do more optimizations, as with ﬁelds). Not all languages support the four styles of visibility used in UML, but Java does.

10.2.4 Accessors It’s a good idea to provide accessor messages for ﬁelds. Accessor messages come in two varieties: getters which return the value of a ﬁeld and setters which set the ﬁeld to a new value (see Implementation Point 5 for an example). Accessors allow us to centralize access Implementation Point 5 This fragment of a Java class shows one ﬁeld with a pair of accessors: ... private int count; public int getCount() { return count; } public void setCount(int c) { count = c; } ...

275

276

Chapter 10

to ﬁelds, making maintenance easier. Accessors are also easy for a compiler to optimize (especially if the associated variable is private). Within the class itself, we may relax the rule slightly: it is common to read a ﬁeld directly but to set it using a setter.

10.2.5 Mapping Classes, Attributes and Compositions In UML, the same notation is used for analysis class diagrams and design class diagrams. However, you will ﬁnd yourself using more of the available notation in your design diagrams. The notation for class ﬁelds is the same as for instance ﬁelds, except that class ﬁelds are underlined. Similarly, class messages use the same notation as instance messages, except that class messages are underlined. You could be forgiven for thinking that this underlining policy is inconsistent with the policy for class names (on a class diagram) versus object labels (on an object diagram) – class names are not underlined while object labels are. Yes, it is confusing. The justiﬁcation is that it’s better to underline whichever version is drawn less often. Figure 10.1 shows a fragment of the analysis class diagram from iCoot being converted into a design class diagram. In mapping to the design diagram, three major issues have been resolved: the classes to be implemented; the attributes’ types; and how to map the composition. In this case, the decision has been made to retain both analysis classes, with no new supporting classes required (apart from the trivial class String). As for the attributes, all of them have survived as private ﬁelds of some appropriate type. An optional attribute

CarModel

CarModel -name:String -price:int

name price

1

1

-details

CarModelDetails

CarModelDetails

engineSize description advert poster

-engineSize:int -description:String -advert:String -poster:String

Figure 10.1: Mapping attributes and compositions from analysis

Mapping the Analysis Class Model into the Design Class Model

would have been mapped in the same way – the only difference would be that, at run time, the corresponding ﬁeld might take a null value. The composition relationship takes a little more thought. For analysis purposes, compositions are bi-directional: starting from either object we can easily ﬁnd the object at the other end. When we come to design, however, we have to deal with the fact that ﬁelds can only point one way (at run time, we can get from the owning object to the object at the far end, but we can’t get back). So, we need to decide whether we want a ﬁeld at each end of the relationship or just at one end (and, if so, which end). We would expect a composition’s primary relationship to be from the composer to the composed. This reﬂects the fact that the composer is the owner of the relationship. It also reﬂects the fact that a composition is similar to an attribute: normally, the owning object uses the services of the attribute or the composed object, but not the other way round. Bear in mind, though, that there are no absolutes: if we felt it suited our particular system better, we could put a ﬁeld inside the composed object or inside both the composer and the composed. (UML also has the notion of a data type, which is designed speciﬁcally for embedded attributes [OMG 03a].) Once we have decided the direction of a composition, we can add an arrow-head to the class diagram to show it, along with a role and a visibility to indicate how the composition maps to a ﬁeld (see Figure 10.1). At run time, messages can only ﬂow in the direction of the arrow. We could also remove the composition altogether and show a -details:CarModelDetails ﬁeld inside CarModel instead. Most UML authoring tools allow the developer to expand and collapse such information at will. In Figure 10.1, the advert and poster ﬁelds are both of type String. This is because we don’t need to manipulate or store raw media in the iCoot system itself. Instead, we can deploy these complex types as ﬁles – *.ram and *.png ﬁles, for example – under the control of a Web server and then store the relative URIs as attribute values. For example, a CarModel could have ‘‘/adverts/mcgs.ram’’ as the value of its advert ﬁeld.

10.2.6 Mapping Other Types of Association Having dealt with compositions, let’s look at the mapping of other types of association. Recall that there are three kinds of association: (plain) association, aggregation (stronger) and composition (stronger still). For mapping purposes, we don’t need to make any distinction between aggregation and association because object-oriented programming languages make no distinction. So, the term association will be used for the rest of this topic. Most of the associations that we decide to retain from analysis, along with any new ones that we add solely for design purposes, end up as ﬁelds on objects. Some may end

277

278

Chapter 10

up as class ﬁelds instead. Either way, since ﬁelds only permit navigation in one direction, we need to decide whether we want to go both ways, or whether one way will do. The way we implement an association depends on the multiplicities at each end: one-to-one, one-to-many or many-to-many.

One-to-One Associations Take a look at the fragment of analysis class diagram at the top of Figure 10.2. Three ways of implementing the association have been shown: we could put a ﬁeld called account into Member, pointing to the InternetAccount; we could put a ﬁeld called member into InternetAccount, pointing to the Member; or we could combine the two, effectively implementing a two-way association. 1

1

Member logs in with

Internet Account

?

1

1 Member

-account

Internet Account 1

1

Member -member -account 1

1

Member -member One-way

Internet Account

Internet Account Two-way

Figure 10.2: Mapping a one-to-one association

If we go for the two-way alternative, we would need to add code to make sure that the association doesn’t get out of step. For example, it would not make sense to change an InternetAccount’s member ﬁeld to point to another Member if we didn’t also inform the original Member that its account had changed (perhaps to null, meaning ‘no value’). It is certainly possible to do this synchronization, using appropriate methods in Member and InternetAccount, but it’s awkward. Therefore, most of the time, we select a one-way alternative. In this case, it’s likely that the natural alternative will work best: let the Member refer to the InternetAccount, but not vice versa. If we choose a one-way alternative, as long as we have a relational database underneath our business layer, we can always derive the complementary information at run time: for

Mapping the Analysis Class Model into the Design Class Model

example, if we choose to put a ﬁeld only into Member, any InternetAccount can, if necessary, ask the database where its Member is. An optional association – with a multiplicity of 0..1 on one end (or on both) – is similar to the one-to-one case, except that the ﬁeld (or ﬁelds) can take the value null (i.e. ‘no object at the far end’).

One-to-Many Associations Figure 10.3 shows an example of a one-to-many association. For this case, we still have to decide whether to put a ﬁeld in one end or in both. The same arguments apply as in the one-to-one case. As before, we may decide that there is a natural solution, but we’re still free to use our skill and judgment as designers to do it another way.

Member

1

*

1..

guaranteed by

Credit Card

?

1

*

1.. Member

-card

Credit Card

*

1.. Member 1

*

1.. Member

-members: List

One-way

Credit Card

1

-members: -card List

Credit Card

non-standard notation

Two-way

Figure 10.3: Mapping a one-to-many association

Unlike the one-to-one case, if we decide to put a ﬁeld in the 1 end this time, we must be prepared to store more than one associated object. For example, if we add a members ﬁeld to CreditCard, the ﬁeld will have to store one or more Member objects. Thus, the CreditCard would need to employ some kind of collection class for holding on to its Member objects (or perhaps an array, according to personal preference). In Figure 10.3, a List has been used. We have something of a dilemma here: the natural choice of collection seems to be Set; however, when you add an object to a Set, most implementations will check whether the object is

279

280

Chapter 10

already there, so that it doesn’t appear twice, potentially an expensive operation. Therefore, if you can be sure that you won’t try to add an object twice, you should use a Bag. If you can’t always be sure, consider using a Bag anyway and performing the duplication test manually in the exceptional cases. If your implementation language doesn’t provide something like a Bag, as Java doesn’t, a List will do, especially if the items are kept in order, which makes for faster searching. The notation shown in Figure 10.3, with regard to the collection-valued ﬁelds, is not quite legal UML. Although we can certainly show a ﬁeld as a navigable association with a role indicating the visibility and ﬁeld name, the :List part is nonstandard. The alternative ways of showing this relationship in UML are more complicated and less informative – the version shown here is compact and it allows us to draw relationships to the type of object inside the collection.

Many-to-Many Associations This is the most complicated case. Consider the example shown in Figure 10.4. Here, we have a Make that can manufacture any number of CarModel objects, while each CarModel can be made by one or more Make objects. Each of the three mapping possibilities is labeled A,

Make

*

1..

makes

*

Car Model

? A Make

*

*

1..

-models: List

Car Model

C

*

1.. Make

B Make

-makes: List

*

1..

*

-makes: List

One-way

* -models: List

Car Model

Two-way

Figure 10.4: Mapping a many-to-many association

Car Model

Mapping the Analysis Class Model into the Design Class Model

B or C – List objects have been chosen again, which, this time, allows us to place the Make objects in order of importance. In common with a lot of many-to-many associations, there is really no owner of the association in this case. If we decide, based on our problem domain, that we’re more likely to start with a Make and then navigate to its CarModel objects, option A would work well. Conversely, if we decide that we’re more likely to start with a CarModel, option B would be better. If, however, we decide that we’re just as likely to start with a Make or a CarModel, we’re really stuck with option C. With option C, the synchronization problem is worse than it was in the one-to-one and one-to-many cases: we’re going to have to search through collections to ﬁnd objects, rather than just accessing individual objects directly. In such cases, it can be worth introducing an association class to deal with the complexity, as described next.

Association Classes We ﬁrst encountered association classes in the analysis section. An association class is necessary if there is data related to an association. The fragment of analysis class diagram at the top of Figure 10.5 shows an example of an association class.

Reservation number:String timestamp:Timestamp state:ReservationState CarModel

Customer

*

* ?

CarModel -model

Reservation -number:String -timestamp:Timestamp -state:ReservationState

Customer -member

Figure 10.5: Mapping an association class

From this diagram, we know that Customer objects can reserve any number of CarModel objects and that CarModel objects can be reserved by any number of Customer objects; but, for each Reservation link, we need to record the reservation number, the time-stamp and the state; this data doesn’t belong to the Customer or to the CarModel, rather it belongs to the

281

282

Chapter 10

association itself. Therefore, we should introduce Reservation as an association class, with appropriate attributes. When we come to designing an association class, the easiest thing to do is to create a design class with ﬁelds for all the attributes, and two extra ﬁelds to point at the associated objects, as shown at the bottom of Figure 10.5. Then, we can create an instance of the association class for each link. If we wanted to be able to navigate from a CarModel to its Reservation objects, we would have to add a ﬁeld to CarModel, with all the associated synchronization problems. (A similar argument applies to Customer.) Alternatively, we can provide a ﬁndAllReservationsFor(:CarModel) message on the ReservationHome class. (A home allows us to create and ﬁnd objects of a particular class, usually by dipping into the underlying database; each home object is a singleton: we write code to guarantee that only one instance of the home exists.) An association class is the most general way of mapping any association from analysis to design. Therefore we could, if we wished, introduce an association class to represent every association from our analysis class diagram. This would alleviate some of the synchronization problems we saw earlier. Although some code generation tools do indeed operate this way, you may feel that it is too much trouble if you’re writing the code by hand. Case Study iCoot BusinessLayer class diagram In Figure 10.6, all the analysis classes of the iCoot BusinessLayer have been mapped into design classes that can be implemented directly in any popular object-oriented language. This class diagram has been enclosed in a UML frame, the name of the package being shown in the top left-hand corner. Details of the ﬁelds for these classes are given in Section B.5.

10.2.7 Universal Identiﬁer Most business objects, at some point in their life, need to be retrieved by key. A key is an attribute value, or a combination of values, that is unique to each instance. For example, a bank account is uniquely identiﬁed by the combination of its account number and sort code. Handling different types of keys in a software system can be cumbersome. Therefore, it’s worth considering the introduction of a unique number to distinguish each business object from other objects of the same class. This helps to synchronize copies of the objects, to track their movements as they travel around a network (their universe) and to handle keys uniformly and efﬁciently (via home objects, for example). Such universal identiﬁers are also

Mapping the Analysis Class Model into the Design Class Model

pkg business

Car Details

-cars:List

-details Car 1

Address

Rental

0..1

1..*

*

*

-address -customer 1

1 1..*

Customer

Member

1 -customer 1

-model CarModel Details

-details

*

1

1 *

*

-card

-model

CarModel

1..*

Reservation *

1

1 Internet Account

Credit Card

1

*

-account

1

-vendor 1 Vendor

1

-category

1..*

Category

Waiting

Needing Renewal

Reservation StateHome

1

-makes:List

Catalog Query

ReservationState

Make

Notifiable

InternetAccount Home

-state

Collectable

Member Home

Car Home

Rental Home

CreditCard Home

Category Home

Vendor Home

Address Home

Displayable

CarModel Home

Reservation Home

CarModel DetailsHome

Concluded

Catalog Query Home

Make Home

CarDetails Home

Figure 10.6: BusinessLayer classes for iCoot

useful when objects and links are stored in a database – the universal identiﬁer takes the place of a memory address when the system is shut down. A suitable type for a universal identiﬁer is an integer. The speciﬁc type of integer depends on how many unique objects of each class you want to be able to accommodate: a 16-bit integer (short in Java) allows 65 thousand unique instances; a 32-bit integer (int in Java)

283

284

Chapter 10

allows four billion unique instances. If four billion proves inadequate, there is always the 64-bit integer ( Java long) which allows a virtually unlimited number of instances (about 18 quintillion). If you do use a universal identiﬁer, it should be set during the construction of the object and it should remain ﬁxed thereafter. Figure 10.7 illustrates the implementation of a universal identiﬁer for the CarModel class. The UML <> keyword has been used to indicate that the CarModel operation creates instances of CarModel. This is good practice in UML since, without the keyword, the reader would have to assume that an operation with the same name as the class is a ‘constructor’ but the fact that constructors exist at all, let alone that their names follow some rule, is language-speciﬁc. (It may be common, but it is still language-speciﬁc.) Therefore, in this book, although the term constructor is used as shorthand for ‘operation used to create instances of the class’, the <> keyword is still used in diagrams. CarModel -id:int -name:String -price:int ~CarModel(id:int) <> +getId():int

Figure 10.7: Implementing a universal identiﬁer

Normally, the constructor would not be public: this would allow the developer to control which universal identiﬁers actually get assigned, using homes for example. (If the constructor were public, any client would be able to create a CarModel using an identiﬁer that was already in use elsewhere.) In Figure 10.7, the constructor has package access, so that the objects can be created by a CarModelHome class living in the same package.

10.3 HANDLING PERSISTENCE WITH A RELATIONAL DATABASE Since most systems have data storage requirements, we’ll look next at how graphs of run-time objects can be mapped into storable data, concentrating on how to use a relational database for this task, since relational databases are by far the most popular in industry. Once we’ve designed our business layer and our database schema, we can think about what code needs to be written to translate from one to the other.

Handling Persistence with a Relational Database

10.3.1 Database Management Systems Often, an application’s data is stored separately, so that the data doesn’t evaporate when the application is shut down – such data is said to persist. A database management system (DBMS) manages arbitrary amounts of data in separate databases – a database is a fenced-off area of data. The terms DBMS and database are often used interchangeably. We use a DBMS to: 1. Invent a schema that describes the data that we’re going to store, using a Data Deﬁnition Language (DDL). 2. Add, remove and update the data in our database using a Data Manipulation Language (DML). 3. Retrieve the data from our database using a Data Query Language (DQL). Ideally, the DDL, DML and DQL should all be declarative. With a declarative language, we specify what we want rather than how to do it. For example, we don’t know, and we don’t want to know, how the DBMS organizes data physically on a disk; instead, we would like to state what kind of data is stored, what the data values should be and what we would like to retrieve. Over the last four decades, many varieties of DBMS have been invented – indexed ﬁle, hierarchical, network, relational and object-oriented, to name a few. There is always a semantic gap (difference in meaning) between the programming language we use to write our code and the way we access data in our database. Therefore, we need to perform some kind of run-time mapping between the software system and the database. Some tools (such as EJB implementations) will generate the mapping code for us. However, even if we wish to use such tools for our project, we still need to understand the underlying principles in order to use them effectively (and in order to understand the run-time error messages). A relational database is often used to store the data in a multi-tier Internet system. The relational model has been chosen here because: • Relational databases are the most common. Although object-oriented databases do exist, they’re much less widely used, especially in the business world. (Most relational databases have object-oriented extensions, but we can ignore those, for the sake of simplicity.) • All relational databases support the same core model, based on rigorous mathematics. This makes them unique among varieties of DBMS. • All commercial relational database implementations support the same hybrid DDL/DML/ DQL language: the Structured Query Language (SQL). • Java has a comprehensive, portable, library for connecting to relational databases, known as Java Database Connectivity ( JDBC).

285

286

Chapter 10

To be fair, the reasons given above for sticking to a relational database are a little optimistic. Some relational DBMSs use nonstandard versions of SQL – always check ﬁrst. The ANSI SQL-92 standard (on which JDBC is based) is also not a particularly rigid speciﬁcation, for example: • With some databases, semi-colons are optional in SQL statements, while in others, they’re compulsory. • You may ﬁnd that your chosen DBMS folds several data types into one: for example, DATE, TIME and TIMESTAMP might all be implemented as TIMESTAMP. • Some data types may not be supported by your DBMS. • The number of bits devoted to a particular data type is not in the standard: an INTEGER might be 16 bits in one DBMS and 32 bits in another. These issues cause some difﬁculties for portable libraries such as JDBC. If you’re not worried about the ﬁner details of accuracy and performance, a relational schema will pretty much work on any of the popular commercial systems, although there may be some porting effort if you switch a completed system from one DBMS to another.

10.3.2 The Relational Model The relational model is a mathematical model, which makes it clean, reliable and easy to optimize (easy for the DBMS, that is). However, the relational model resembles a ﬁling cabinet containing indexed cards, rather than a repository of complex, highly-connected objects. So, the semantic gap is large. Although an object-oriented model can be mapped easily onto a relational schema, it’s difﬁcult to predict which of several possible mappings will be the most efﬁcient for any particular software system (with respect to storage efﬁciency and access speed). Therefore, you may need to experiment, preferably with the help of a mapping tool or a persistence layer, to ﬁnd which mapping works best for your system. Here, we’ll look at one possible mapping: a straightforward, pure mapping. This way, you should be convinced that it’s possible to store object-oriented data in a relational database. For this mapping, we can assume that we have control over how the data is stored. In real life, we may not be so lucky: the schema may be under the control of Database Administrators (DBAs) or the database may already exist as a legacy system that we can’t modify – in these cases, we would have to ‘reverse engineer’ our mapping. We won’t look at how to write the code to ship data between a software system and its database, because that is probably the single most difﬁcult piece of code in the whole system. If you use a framework such as EJB, the tools will generate all of the code you need for the basic mapping described here. To do anything manually requires lots of effort and

Handling Persistence with a Relational Database

study of your database’s programming interface. For our purposes, it’s enough to say that the underlying techniques involve using SQL statements (via a library such as JDBC) to read data from the database into run-time objects and then to write new data back out. Since the database layer is encapsulated by the business layer, we’re free to tune the business layer for the beneﬁt of client programmers and the database layer for the beneﬁt of the database – the mapping code can do the rest.

Tables The relational model is based on tables of data (also known as relations) which contain columns and rows. Figure 10.8 shows a table called ADDRESS, with four columns of string data (HOUSE, STREET, COUNTY and POSTCODE) and one of integer data (ID). As you can see, the ADDRESS table has three separate addresses stored in it. Each column stores the values for one of the address attributes – thus, the address with ID 2 represents ‘8 Yewbrook Rd, Cheshire, SK4 3QT’. ADDRESS ID HOUSE 2 8 9 Dunroamin 6 74

STREET Yewbrook Road Dairy Avenue Old Ladbroke Grove

COUNTY Cheshire Greater Manchester Lancashire

POSTCODE SK4 3QT M19 4IK M20 7HJ

Figure 10.8: An ADDRESS table

Each column can store values of one particular type. The SQL standard deﬁnes a couple of dozen types that we can choose from. For the purposes of this book, we can make do with the following: • • • • •

VARCHAR(X):

A string, up to a maximum of X characters. INTEGER: A whole number which is often, but not always, 32 bits. DATE: A day in the Gregorian calendar. TIMESTAMP: A combination of date and time of day. BOOLEAN: True or false.

For example, for the ADDRESS table above, we might specify INTEGER, VARCHAR(20), and VARCHAR(10) as the types of ID, HOUSE, STREET, COUNTY and POSTCODE, respectively. With the relational model, every row must be unique. For the ADDRESS table, this is achieved by introducing a unique ID attribute. Since every row is unique, each table is actually a set of rows. The rows in Figure 10.8 could have been shown in any order. VARCHAR(40), VARCHAR(40)

287

288

Chapter 10

Keys A key is a value, or a combination of values, that uniquely identiﬁes a row. Some tables contain obvious keys: for example, an insurance policy always has a policy number. Other tables need to have keys invented for them: for example, a customer table would need to include a customer number to identify unique people who buy our products. A key that combines several values (such as a bank account’s number and sort code) is referred to as a compound key. Some tables have more than one candidate key: for example, a car has a license number as well as a Vehicle Identiﬁcation Number (VIN) that appears on a plate riveted to its body. In such cases, we have to choose one candidate to act as the primary key – the DBMS will assume that the primary key is the one most often used to ﬁnd a particular row and it will perform optimizations accordingly. Finding a row using a compound key is slower than ﬁnding a row using a simple key. For example, considering the ADDRESS table, we could use (HOUSE and POSTCODE) as the key, because the combination is unique for every house. However, a combination of strings is inconvenient and slow to use. Thus, in the ADDRESS example, an artiﬁcial key (ID) has been introduced that uniquely identiﬁes each dwelling. This also allows us to record the fact that two customers live at the same address (we could store the ID twice, but the address only once). On the schema diagrams in this book, primary keys are shown in bold. If you are sketching such a schema by hand, you can add a ‘+’ next to the column name instead.

Mapping an Object Model to a Relational Model When mapping an object model into tables, we can start with either the analysis class diagram or the design class diagram. The analysis diagram is closer to a relational model, because it shows no commitment to the direction of associations, so it would seem to be a good choice. However, the design class diagram has types assigned to ﬁelds, which we need to know when designing the tables. To solve this paradox, we will use the analysis model to drive the design of the tables, but pick up the types from the design model. Any discrepancy between the actual design model and the database tables can be removed by the mapping code that we write.

10.3.3 Mapping Entity Classes In order to map an entity (business object) from an object-oriented model into a relational schema, we need to introduce a table with the same name as the entity’s class. (We’ve already seen an example of this with the ADDRESS table.) Every row in an entity table represents a unique object from the business domain. For each simple ﬁeld (primitive or string) we can add a column to the table with the same name as the ﬁeld and an appropriate SQL data type. Entity ﬁelds that point to (non-string)

Handling Persistence with a Relational Database

objects must be treated differently, as we will see shortly when we look at the mapping of associations. For the sake of object-oriented programming, it’s important to introduce an integer attribute, such as ID, to use as the primary key. This is because it simpliﬁes the mapping code (important if we’re writing the code manually) and it makes navigating from one object to another easier and more efﬁcient (especially important since objects tend to form complex graphs). It also provides the beneﬁts associated with universal identiﬁers. From here on, an INTEGER column called ID will be added to every entity table, even though it is not shown in the accompanying class diagram fragments.

10.3.4 Mapping Associations When we looked at how to map an analysis class model into a design class model, we saw that we had to transform the two-way analysis associations into one-way pointers. A relational database stores two-way associations directly, so we don’t have the same problem: if our database schema allows us to navigate from entity A to entity B, the DBMS and the query language will allow us to navigate just as easily from entity B to entity A. Therefore, our database schema will resemble the analysis class model more closely than the design class model. (We should still design the business layer before producing the database schema though, because the design process will harden the analysis model and help us to choose attribute types.)

One-to-One Associations For a one-to-one association, we can add a foreign key to one of the entity tables. A foreign key is an entry in one table that refers to a primary key in another table. In other words, a foreign key is a reference from a row in one table to a row in another table. To emphasize which columns represent foreign keys, the column names are shown in italics here. (When drawing by hand, you can add a ‘>’ next to the column name instead.) In Figure 10.9, the CARMODEL table contains a foreign key called CARMODELDETAILSID that allows us to ﬁnd a CarModel’s details in the CARMODELDETAILS table. (The fragment of class diagram shown here uses analysis notation, without navigability, but universal identiﬁers have been added to each table, even though they’re not analysis-level attributes.) We could equally well add a foreign key called CARMODELID to the CARMODELDETAILS table – the choice made reﬂects the logical nature of the composition relationship that we’re dealing with. Because of the bi-directional properties of relational databases, we don’t need to put a foreign key in both ends. As an alternative to foreign keys, we could combine the two tables into one. For ease of maintenance, tables should only be combined when one of the tables doesn’t represent an entity in its own right. For example, if CarModelDetails were just a composite attribute of CarModel, budded off for the convenience of programmers, it would be reasonable to add the

289

290

Chapter 10

CarModelDetails CarModel name:String price:int

1 description:String engineSize:String advert:String poster:String

CARMODEL NAME ID 111 Grey Shadow Fly 39 Dooby 9 14 CARMODELDETAILS ID ENGINESIZE 19 3500 18 3000 37 4800

PRICE 19500 25000 35000

DESCRIPTION Pure luxury... Power and... Smooth but...

CARMODELDETAILSID 37 19 18

ADVERT arf.ram amd9.ram rcgs.ram

POSTER arf.jpg amd9.jpg rcgs.jpg

Figure 10.9: Mapping a one-to-one association to a foreign key RELEASEDATE, CATALOGDATE, CLIP

and POSTER columns to the CARMODEL table and discard CARMODELDETAILS altogether. In this case, we can tell that the developers have decided to regard CarModelDetails as an entity because it has been given a universal identiﬁer. For the special case of an optional association (multiplicities of 1 and 0..1), we can add a foreign key to the optional end. Incidentally, relational databases support nullable columns. A nullable column is one that allows a cell to contain the value NULL, meaning ‘there is no value here’. Thus, we could also store an optional association as a nullable foreign key in the 1 end. However, for reasons of simplicity and elegance, it’s better to avoid nullable values.

One-to-Many Associations For a one-to-many association, we can add a foreign key to the ‘many’ table, as shown in Figure 10.10. In this example, each CreditCard can guarantee more than one Member. Thus, each row in the MEMBER table has a reference to its card (CARDID). (This assumes that each member only has one card recorded in the system.) In the ﬁnished system, this MEMBER table also has a foreign key to indicate the member’s address.

Many-to-Many Associations For many-to-many associations, one foreign key is not enough to identify the many entities at each end of the association. Consider the fragment of analysis class diagram shown at the

Handling Persistence with a Relational Database

Card

Member number:String inGoodStanding:bollean

MEMBER ID NUMBER 4 M105 11 M9371 2 M203

CARDID 14 45 14

CARD ID TYPE 14 Visor 45 Annex

*

number:String 1 type:String expiryDate:Date

INGOODSTANDING TRUE FALSE TRUE

NUMBER 1111 2222 3333 4444 7777 8888 9999 5555

EXPIRYDATE 2006-10-09 2006-12-14

Figure 10.10: Mapping one-to-many associations to foreign keys

top of Figure 10.11. Here, a Make can manufacture many CarModel objects and a CarModel can be manufactured by many Makes (the Make objects are collaborating in the manufacture of the CarModel). In the pure relational model, every value in a table must be atomic – i.e. not a collection of values. So, we wouldn’t be able to record all of a CarModel’s Make objects as a MAKEIDS column in the CARMODEL table. Since we can’t have multi-valued attributes, we need to use a link table. As its name suggests, each row in a link table represents a link between an entity in one table and an entity in another table. The bottom half of Figure 10.11 shows a link table called MAKECARMODEL that links CarModel objects to Make objects. For example, we can see that Make 8 and Make 9 are both manufacturers of CarModel 39, and that Make 8 also manufactures CarModel 111. Technically, a link table has a compound primary key consisting of two foreign keys. We could also use a link table to store one-to-one and one-to-many associations. However, they’re probably best left for the special cases of many-to-many associations and association classes (see next).

Association Classes Association classes, since they have data of their own, must be mapped to link tables, regardless of the multiplicities at each end of the association. Unlike plain link tables, tables representing association classes have attribute columns – they may even have an ID column (if the association class represents an entity in its own right).

291

292

Chapter 10

CarModel

Make name:String 1..

*

*

1..

CARMODEL ID NAME 111 Grey Shadow Lacrosse 39 14 Dooby 9

MAKECARMODEL CARMODELID MAKEID 8 111 65 14 9 39 8 39

name:String price:int

PRICE 19500 25000 35000

CARMODELDETAILSID 37 19 18

MAKE ID NAME 65 Astra Marten 9 Alpha Rodeo 8 Rolls Choice

Figure 10.11: Mapping a many-to-many association to a link table

For example, Figure 10.12 shows a table representing the association class called Reservation. The RESERVATION table has a primary key for the reservation itself, two foreign keys referencing the objects at each end of the association and two attribute columns.

10.3.5 Mapping Object State For business objects that have an associated state machine, such as might be shown on a state machine diagram, we need to record the state that each object is in. In the business layer, our business object might have a simple ﬁeld indicating its state, a String or an int perhaps, or we might have a complex ﬁeld pointing to a state object (as described by the State pattern). If we choose to use a String or an int, we need to make sure that the ﬁeld can only take on certain values: for iCoot, Reservation has six possible states, so we would have to choose six Strings (‘‘waiting’’, ‘‘notiﬁable’’, ‘‘collectable’’, ‘‘needingRenewal’’, ‘‘storable’’ and ‘‘concluded’’, for example) or six integers (0 to 5 perhaps). From the database point of view, we could also take a simple approach: add a column called STATE to the RESERVATION table and set its type to VARCHAR or INTEGER, as appropriate. (We would expect the INTEGER version to be faster, but harder to debug.) This approach is illustrated in Figure 10.13. Or, we could take a more complex approach: introduce a new table for each possible state and use foreign keys to indicate which reservations are in which states. For example, Figure 10.14 shows two of the six state tables that we would need, indicating that reservations

Handling Persistence with a Relational Database

Reservation number:String timestamp:Timestamp CarModel

*

*

Customer

RESERVATION ID 7 1 99

CARMODELID 33 22 72

CUSTOMERID 4 2 4

NUMBER R187a R7b R459b

TIMESTAMP 2004-12-06 14:23:16.543 2004-12-03 00:03:21.872 2004-12-05 09:45:07.210

Figure 10.12: Mapping an association class to a link table

RESERVATION CARMODELID ID 7 33 1 22 99 72

CUSTOMERID 4 2 4

NUMBER R187a R7b R459b

TIMESTAMP 2004-12-06 14:23:16.543 2004-12-03 00:03:21.872 2004-12-05 09:45:07.210

Figure 10.13: Mapping object state to a column

WAITINGRESERVATION RESERVATIONID 7 99 COLLECTABLERESERVATION RESERVATIONID 1

Figure 10.14: Mapping object state to state tables

STATE 0 2 0

293

294

Chapter 10

7 and 99 are waiting and reservation 1 is collectable. The complex approach is arguably more

elegant. Whether or not it’s more efﬁcient than the simple approach depends on the usage patterns within our system, a difﬁcult thing to predict. When storing states, we sometimes have to deal with state attributes – data associated with a business object when it’s in a particular state. For example, when a reservation is in the waiting state, we need to know when the reservation was last renewed (because, after a week in the waiting state, the reservation needs to be renewed). Similarly, for a collectable reservation, we need to know when the car was made collectable (because, if the customer doesn’t collect it within three days, the car becomes storable). With the state-table approach, we can store state attributes as extra columns. This is illustrated, for our two example states, in Figure 10.15. With the state-column approach, we would have to add a column for every state attribute to the business object table, which could be many columns. Each column would have to be nullable so that the data could be omitted when the object was not in the relevant state. This approach is illustrated for our two example states in Figure 10.16. Since it’s better to avoid nullable columns, the state table approach wins in this case. WAITINGRESERVATION RESERVATIONID 7 99

LASTRENEWEDDATE 2004-16-10 2004-10-18

COLLECTABLERESERVATION DATEPUTASIDE RESERVATIONID 2004-12-04 1

Figure 10.15: Adding state attributes to state tables RESERVATION ID CARMODELID 7 33 1 22 99 72

...

CUSTOMERID 4 2 4

STATE 0 2 0

NUMBER R187a R7b R459b

TIMESTAMP 2004-12-06 14:23:16.543 2004-12-03 00:03:21.872 2004-12-05 09:45:07.210

LASTRENEWEDDATE 2004-16-10 NULL 2004-10-18

DATEPUTASIDE NULL 2004-12-04 NULL

Figure 10.16: Mapping state attributes to nullable columns

...

Handling Persistence with a Relational Database

Mapping Inheritance In order to map an inheritance hierarchy onto tables, we can introduce a table for each class, with columns for the attributes added by that class. So that we can ﬁnd all the attributes for an object, the tables must share the same primary key. For example, Figure 10.17 shows the Customer class hierarchy mapped to three tables: the CUSTOMER table has three columns for the attributes deﬁned by the Customer class; the NONMEMBER table has one column for the attribute deﬁned by the NonMember class; and the MEMBER table has two columns for the attributes deﬁned by the Member class. NonMember driversLicense: String

Customer name:String phone:String amountDue:int

Member inGoodStanding: boolean number:String

CUSTOMER ID NAME 4 Helen Meeder 11 Alice Tara 2 George Milo

AMOUNTDUE 0 0 3980

PHONE (0452)94983 (0161)898349 (07968)14599

33

0

(07968)14599

Mary Smith

MEMBER ID 4 11 2

NUMBER M105 M9371 M203

INGOODSTANDING TRUE FALSE FALSE

CARDID 14 45 14

ADDRESSID 9 2 9

NONMEMBER ID 33

DRIVERSLICENSE DUCK MH8L3 7324

Figure 10.17: Mapping inheritance to multiple tables

As usual, the universal identiﬁer (which will have been added to Customer class) is used as the primary key: this time it appears as the primary key in all three tables. When we come to retrieve the data for Customer 33, for example, we pick up NAME, AMOUNTDUE and PHONE from the CUSTOMER table; next, we have to test whether there is a row with key 33 in either of the subclass tables (if there is, we know that Customer 33 is actually an instance of

295

296

Chapter 10

a subclass); in our example, we detect that Customer 33 has NONMEMBER data, so we pick up its License. The end result is the NonMember shown in Figure 10.18.

aNonMember

id = 33 name = "Mary Smith" phone = "(07968) 145 99" amountDue = 0 driversLicense = "DUCK MH8L3 7324"

Figure 10.18: A reconstituted nonmember

As with the mapping of object states, we could instead put all the attributes into a single table (this would reﬂect the fact that an object being an instance of a class is similar to an object being in a particular state). For example, in Figure 10.19 we have a single CUSTOMER table with a CLASS column, specifying the class of each individual row. In addition, the table has a nullable column for each of the attributes in the hierarchy. As ever, the choice of mapping to multiple tables or a single table will be based on personal preference, experience, crystal-ball gazing and experimentation. CUSTOMER ID NAME 4 Helen Meeder 11 Alice Tara 2 George Milo 33 Mary Smith

AMOUNTDUE 0 0 3980 0

...

NUMBER M105 M9371 M203 NULL

INGOODSTANDING TRUE FALSE FALSE NULL

...

DRIVERSLICENSE NULL NULL NULL DUCK MH8L3 7324

PHONE (0452)94983 (0161)898349 (07968)14599 (07968)14599

CARDID 14 45 14 NULL

...

ADDRESSID 9 2 9 NULL

CLASS 0 0 0 1

Figure 10.19: Mapping inheritance to a single table

...

Handling Persistence with a Relational Database

ADDRESS (ID:INTEGER,HOUSE:VARCHAR(99),STREET:VARCHAR(99),COUNTY:VARCHAR(99), POSTCODE:VARCHAR(99)) CAR (ID:INTEGER,TRAVELLED:INTEGER,DATELOST:DATE,CARDETAILSID:INTEGER) CARD (ID:INTEGER,TYPE:VARCHAR(99),NUMBER:VARCHAR(99)) CARDETAILS (ID:INTEGER,BARCODE:VARCHAR(99),NUMBERPLATE:VARCHAR(99),VIN:VARCHAR(99)) CARMODEL (ID:INTEGER,NAME:VARCHAR(99),PRICE:INTEGER,CARMODELDETAILSID:INTEGER, CATEGORYID:INTEGER,VENDORID:INTEGER) CARMODELDETAILS (ID:INTEGER,ENGINESIZE:VARCHAR(99),DESCRIPTION:VARCHAR(256), ADVERT:VARCHAR(99),POSTER:VARCHAR(99)) CATEGORY (ID:INTEGER,NAME:VARCHAR(99)) COLLECTABLERESERVATION (RESERVATIONID:INTEGER,DATENOTIFIED:DATE) CONCLUDEDRESERVATION (RESERVATIONID:INTEGER,REASON:VARCHAR(99)) CUSTOMER (ID:INTEGER,NAME:VARCHAR(99),PHONE:VARCHAR(99),AMOUNTDUE:INTEGER) DISPLAYABLERESERVATION (RESERVATIONID:INTEGER,REASON:VARCHAR(99)) INTERNETACCOUNT (ID:INTEGER,PASSWORD:VARCHAR(99),SESSIONID:INTEGER) MAKE (ID:INTEGER,NAME:VARCHAR(99)) MAKECARMODEL (CARMODELID:INTEGER,MAKEID:INTEGER) MEMBER (ID:INTEGER,NUMBER:VARCHAR(99),INGOODSTANDING:BOOLEAN,CARDID:INTEGER, ADDRESSID:INTEGER) NEEDINGRENEWALRESERVATION (RESERVATIONID:INTEGER,DATERENEWALNEEDED:DATE) NONMEMBER (ID:INTEGER,DRIVERSLICENSE:VARCHAR(99)) NOTIFIABLERESERVATION (RESERVATIONID:INTEGER,DATEPUTASIDE:DATE) RENTAL (ID:INTEGER,NUMBER:VARCHAR(99),STARTDATE:DATE,DUEDATE:DATE,TOTALAMOUNT:INTEGER) RENTALCAR (RENTALID:INTEGER,CARID:INTEGER) RESERVATION (ID:INTEGER,NUMBER:VARCHAR(99),TIMESTAMP:TIMESTAMP,CUSTOMERID:INTEGER, CARMODELID:INTEGER) VENDOR (ID:INTEGER,NAME:VARCHAR(99)) WAITINGRESERVATION (RESERVATIONID:INTEGER,LASTRENEWEDDDATE:DATE)

Figure 10.20: Database schema for iCoot

297

298

Chapter 10

Case Study iCoot database schema The ﬁnished database schema for iCoot is shown in Figure 10.20. Each table is shown as a name followed by column types in parentheses. As before, primary keys are shown in bold and foreign keys are shown in italics (with primary–foreign keys in bold–italic). For completeness, this schema includes a few pieces from the full Coot system, such as NONMEMBER and DATELOST.

10.4 FINALIZING THE USER INTERFACES Next, we’ll look at the design of our user interfaces. We have a handful of vague sketches that we used while ﬂeshing out the system requirements. This section includes hints and tips on how to produce good, simple, interfaces for thin clients. In the early stages of development, during requirements capture and analysis, it’s useful to consider the functionality of the user interfaces. This makes sense because, with a usecase-driven methodology, the way that actors interact with the system is paramount and most actors are human. To this end, we already have the following: • User interface sketches. These were used to help us produce system use cases during requirements capture, with the help of our sponsors. • Boundary objects in communication diagrams. During dynamic analysis, we used communication diagrams to show the realization of the use cases; in these diagrams, every actor was shown interacting with the system via boundary objects. But, we still need to design the user interfaces: we have to take the coarse boundary objects, the vague user interface sketches and the precise use cases, and transform them into user interface descriptions that can be implemented directly. (We expect system-to-system interfaces to be designed as part of our layers – for systems that are accessed by us – or as part of our business services – for systems that access our system.) You may have been surprised by the fact that most of a software system can be designed successfully without considering the user interfaces in detail (having completed the system use cases, we concentrated almost entirely on the objects inside the system). There are two main reasons for this: • The correct behavior of a system depends on its internal construction, not on the way that people interact with it. To use an analogy: a car consists of an engine, four wheels and a

Finalizing the User Interfaces

body; depending on which interfaces we use, and how we use them, we can get a car to transport people from A to B, or we can get it to tow a caravan, or we can get it to knock down a wall; however we choose to interact with a car, it still behaves as a car. • We would prefer to write reusable code. If we focus on the needs of a particular set of interfaces, we risk producing a system that only works for those interfaces – this is one of the problems of traditional development: ‘Solve today’s problem today, forget about tomorrow’. Being use-case driven might seem to be against the principle of reuse; however, as far as the internals of the system are concerned, use cases are just a good way of ensuring that the developers don’t wander off into irrelevant areas. Rather than delving into the theory of human–computer interaction, we’ll look at some basic principles for good user interface design (especially for thin clients accessing multi-tier systems).

Be Guided by Use Cases From the system point of view, use cases simply keep the developers on track. To the users, on the other hand, the use cases are everything. Thus, use cases should be used to give structure to the user interfaces. In general, we should try to keep our use cases simple: they should not contain too much functionality or they will be difﬁcult to manage. Therefore, although we would expect each user interface to represent a number of use cases, the interfaces themselves can still be simple. We would expect the grouping of related use cases to be reﬂected in the structure of the user interface. For example, iCoot customers are presented with a single, Web-based, user interface that includes a dozen or so use cases. Within that single interface, we would expect to ﬁnd activities grouped into ‘reservations’, ‘rentals’, ‘browsing the catalog’ and so on. We should also avoid splitting a single use case, or a chain of related use cases, into more than one interface. For example, a customer who had homed in on an interesting car model would be annoyed if they had to go to a different interface to reserve it.

Keep it Simple Keeping things simple is a good principle in its own right. But also, many of our users will be novices, especially those users accessing our system over the Internet, so simplicity is even more important. There’s a strong argument for keeping user interfaces simple and uncluttered for expert users too: we don’t want the overwhelming functionality of our user interface to hamper their productivity. In an e-commerce context, we don’t want a learning barrier in front of the goods that we have to sell. When a customer comes across our site, if they’re presented with long instructions on how to use it, they’ll leave immediately. So that customers can use our site

299

300

Chapter 10

immediately, we must limit ourselves to trivial, step-by-step, guidance, as in Click here to buy or Enter your credit card details below and click Next. The simpler the interface, the easier it will be to port to more primitive platforms (for example, mobile phones, set-top boxes, personal digital assistants, home appliances). The need to produce portable user interfaces is a strong argument for using a control layer in your design.

Use Notebooks You should use notebooks to group related use cases. A notebook is a set of pages that shows only one page at a time – the other pages are available via tabs along the edge. The advantage of a notebook is that, because its size and location stays the same, the user can focus on a single area of the screen, even for long interactions that involve more than one use case. This improves the user experience and also user productivity. Notebooks are supported by most GUI libraries, such as Java’s Swing (Figure 10.21 shows iCoot running as a Swing applet). For HTML/CGI interfaces, we can only simulate a notebook – the end result is similar, but

Figure 10.21: iCoot using a Swing notebook

Finalizing the User Interfaces

the user has to tolerate the ﬂashing-to-background-color and longer delays when moving to a new page (Figure 10.22 shows iCoot running as a simulated HTML/CGI notebook).

Figure 10.22: iCoot using a simulated notebook

Often, a use case will map cleanly to a single page. With the help of a wizard, even a complex use case can be followed inside a single panel.

Use Wizards A wizard is a sequence of pages that guides a user, step by step, through a complex activity. When the user selects an activity, they’re presented with a page representing step 1; the user completes step 1 and clicks the Next button to go on to step 2; and so on. Before completing an activity, a user can revisit the earlier steps, using the Back button. Each page of the wizard contains ﬁelds, lists and other widgets that the user employs to enter the data needed for the current step. Normally, each page will have simple instructions, such as Type your old password into the box below and then click Next.

301

302

Chapter 10

The step-by-step nature of the wizard and the simple instructions allow the user to perform a complex activity without having to remember how to do it. Also, a wizard, like a notebook, keeps the user’s focus on one area of the screen. As with notebooks, if we’re using a pure HTML/CGI interface, we can only simulate wizards, but the end result is probably better than if we didn’t use them at all. (In this case, the alternative would be one very long HTML page containing all the steps of the activity: rather inconvenient for the user to have to scroll down through the steps.)

Avoid Multiple Windows Multiple, overlapping windows were invented for the beneﬁt of computer experts rather than novice users. Even if our user is sufﬁciently computer-literate to open a Web browser, we should not expect them to be able to cope with multiple browser windows or pop-up dialogs. Apart from the inherent complexity, we also have to accept that, on some platforms, multiple windows just don’t make sense. In iCoot, the user navigates to a particular car

Wrong Figure 10.23: Problems with multiple windows

Finalizing the User Interfaces

model and clicks the Make a Reservation button to reserve it. Since there is a potential ﬁne associated with reserving a car model, the system needs to ask the user for conﬁrmation. The natural thing for a developer of multi-windowed systems to do is to pop up a dialog. However, iCoot may be running on a TV (with the help of a set-top box). Due to limited screen resolution and space, set-top boxes don’t support dialogs, so there’s nowhere for the message to go (see Figure 10.23). It is possible to avoid multiple windows (see Figure 10.24) by overlaying the previous page with the contents of the dialog. If the user clicks No, the system returns to the previous page (equivalent to canceling the dialog); if the user clicks Yes, the reservation is made and the system returns to the previous page.

Right Figure 10.24: Avoiding multiple windows

The interface shown in Figure 10.24 is a wizard running inside a notebook: the radio buttons down the left-hand side of the screen correspond to the tabs of the notebook, allowing the user to switch between major use cases; within the notebook page, we see a wizard guiding the user through the U7:Make Reservation use case (which is an extension of U1:Browse the Index).

303

304

Chapter 10

So, notebooks and wizards can be used together to conﬁne complex interactions to a single area of the screen, so that the user doesn’t get tired or confused. This all works well as long as we stick to the ‘Do one thing at a time’ rule; even when we design our user interface according this rule, we can still allow the user to switch between half-ﬁnished activities. For example, in Figure 10.24, the user can ignore the question for the time being and switch to the Change PIN page. When the user eventually returns to the Index page, the same question will be waiting to be answered.

10.5 DESIGNING THE BUSINESS SERVICES Once we’ve decided on the look and feel of the user interface and the interface of our business layer, we can design our server layer: this consists of a number of server objects, collaborating with objects in the business layer, providing a simple interface for the beneﬁt of clients. Business services are the queries and commands that a middle tier makes available to its clients. For example, for iCoot, we might derive the following set of business services: • • • • • • • • • • • • •

Read the index headings from the Catalog. Read CarModel for a given index heading. Read every CarModel Category. Read all CarModel engine sizes. Read all Makes of CarModel. Read CarModels for a given set of Category, engine size and Make. Read details for a given CarModel. Reserve a CarModel. Read details for a given Member. Change a Member’s password. Read the Cars rented by a given Member. Read the Reservations made by a given Member. Cancel a Reservation.

The business services represent a concise summary of the information ﬂow between the user interfaces on the client and the business logic on the middle tier. Business services can be derived by walking through the use cases, with reference to the user interface design and the system architecture. (The idea of breaking a server’s interface down into the services it provides can be applied to any kind of communication; for example, we could extend the ideas discussed in this section to the problem of deciding how the layers within a subsystem should communicate.)

Designing the Business Services

For business services to make sense in the context of HTML/CGI interfaces, we have to assume that user interfaces are spread between client tier and middle tier: the combination of HTML forms, servlets and JSP result pages comprise a single user interface, despite the fact that some pieces run on the middle tier. We can even execute servlets on a separate middle tier machine to the one that’s performing the business processing, to reduce the load on the business objects – in this case, the servlet machine is really a client of the middle tier rather than part of the middle tier itself. Therefore, in the following sections, the term client, means either: • A piece of code, such as an applet, running on a separate machine to the business layer. • Servlets running in a separate process or on a separate machine to the business layer. Implementing client–server communication in terms of business services allows us to: • Simplify the client code. Since most client interfaces only need a subset of the system’s capability, each one can be given a cut-down view. • Complicate the business layer. We can make the business layer as complex, powerful or reusable as we like, or even change it altogether, without worrying about the impact on the client. • Build a pluggable server layer. We would like to provide various kinds of interface to cope with different user preferences or client capabilities – applet/RMI versus HTML/CGIwith-servlets, for example – without having to re-implement the middle tier services every time.

10.5.1 Using Proxies and Copies When a client requests a business service, the result may include one or more business objects. For example, if the client request is ‘Read all Make objects of CarModel’, the result will be a list of makes. As well as retrieving business objects, a client may send business objects to the server. For example, if the client request is ‘Cancel a Reservation’, the reservation will need to be speciﬁed. So, logically at least, business objects ﬂow back and forth between client and server. If we were dealing with a one-tier system, there would be no problem: we would just pass object pointers around our run-time system and send messages to the objects according to our needs. However, when we’re dealing with a network, things are not that simple. We can, with the help of technologies such as RMI, arrange for messages to pass between objects over the network: each object lives inside a single run-time system, but messages can arrive from outside. Alternatively, we can pass copies of objects around the network: each run-time system has a copy of the objects it needs and sends messages to them locally. In the case of communication between layers in a subsystem, we have a similar choice: do we pass

305

306

Chapter 10

object references between layers or object copies? (We would expect an open layer to pass references across its boundary while a closed layer might pass copies.) To be more precise, we have two basic strategies to choose from: proxies are client objects that know how to forward the messages they receive to a real business object sitting elsewhere; copies are client objects containing a copy of the real business object’s data. The main advantages of using proxies are that: • All clients see the same objects, so they’re always working with the same information. • All the run-time systems merge into a single universe: distributed objects are treated in the same way as local objects. The main disadvantages of proxies are that: • The business layer objects must be secured. For example, we must be sure, when we receive a message from a client, that the client has permission to send the message. This complicates the business layer. (The alternative security model, where the business objects are accessed only by the server layer and the server layer only has to secure its business services, is much easier to implement.) • Network trafﬁc increases. In an object-oriented system, we normally get hold of the object and then send it messages; with the proxy technique, initial retrieval of the object is quick but sending messages to the object is slow (because each message must travel over the network). • The burden on the middle tier increases: the business layer, running on the middle tier, executes all of the methods all of the time. The main advantages of using copies are that: • Network trafﬁc reduces: we retrieve more information to begin with (all the object’s data) but methods execute quickly (because they run locally). • Processing doesn’t all take place in the middle tier, because some business methods execute on the client. • Direct access to business methods doesn’t have to be secured because only the server layer has direct access and the server layer is secure in itself. • The object’s implementation doesn’t have to be concurrent-safe: once a copy has been created on behalf of a client, the client has exclusive access to it. The main disadvantages of copies are that: • Too much data may be copied. For example, if we copy a CarModel, we might end up copying all of its CarModelDetail objects, all of its Make objects, all of its Reservation objects, all of the Member objects that have made those reservations, and so on.

Designing the Business Services

• The copies can get out of step because each client has an independent copy of every object that it uses. For some applications, it may be quite reasonable for a client copy to become out of date. For example, if a user issues a search engine request from a Web browser, the results page that is displayed will age as Web pages are added, updated and removed. The user has to accept that, strictly speaking, the results were only accurate at the moment the query was executed. By careful coding, we can alleviate some of the problems of proxies and copies. For example, when using proxies, we could cache some of the object’s data locally to reduce network trafﬁc, taking care to ensure, by some complicated mechanism, that the objects didn’t get out of step. With the copying approach, we can copy any referenced objects on demand, thereby reducing the copy-too-much problem: for example, when we read a CarModel, the client can be given the top-level attributes only – any referenced objects would stay as null unless the client actually navigated to them (using encapsulation, this process could be made transparent to the client code). If we’re lucky, we will have access to a framework or a library that supports conﬁgurable proxies, conﬁgurable copies and hybrids of the two. However, even with such technologies, the developer still has to choose which variety they want to use. Alternatively, as long as we’re prepared to accept a little hand-crafting, we can take a simpler approach: lightweight copies. With this approach, when a client asks for a business object, it is given only the essential information it needs and information is not passed to the client if the client must already have that information. Going the other way, when a client needs to identify a business object to the server, the client passes just the universal identiﬁer of the object to the server. For lightweight copies to work, we need business services (so that we can reason about what information is needed by the client) and universal identiﬁers (so that clients have an efﬁcient way of passing objects back). The information passed to a client will consist of copies of the business object’s simple attributes (numbers, strings, and so on) and universal identiﬁers for any reference attributes, in case the client needs to navigate further.

10.5.2 Classifying Business Services We would expect business services to form groups of related behavior, in other words, messages on objects. We would expect these server objects to have few attributes of their own and we wouldn’t expect them to record state on behalf of the clients (each client must remember its own state). Server objects reside in the server layer of a multi-tier system. The server objects that have been designed for iCoot are shown in Figure 10.25. The business services have been classiﬁed as reservations, authentication, membership information,

307

308

Chapter 10

CatalogServer +readCategoryNames():String[] +readMakeNames():String[] +readEngineSizes():int[] +readIndexHeadings():String[] +readCarModels(q:PCatalogQuery):PCarModel[] +readCarModels(heading:String):PCarModel[] +readCarModelDetails(carModelId:int):PCarModelDetails

AuthenticationServer +logon(membershipNumber:String,password:String,steal:boolean):long +logoff(sessionId:long)

MembershipServer +readMember(sessionId:long):PMember +changePassword(sessionId:long,old:String,new:String)

RentalsServer readRentals(sessionId:long):PCar[]

ReservationsServer +readReservations(sessionId:long):PReservations[] +createReservation(sessionId:long,movieId:int) +deleteReservation(sessionId:long,reservationId:int)

Figure 10.25: Server objects for iCoot

catalog information and rentals (ReservationsServer, AuthenticationServer, MembershipServer, CatalogServer and RentalsServer, respectively). The business service messages shown in Figure 10.25 have been tuned to ensure that neither the client nor the server will pass information to the other side unnecessarily. For example, consider the CatalogServer. In order to search for car models, the client must ﬁrst retrieve all the category names, engine sizes and make names. These come back from the server as String objects and int objects. The client then allows the user to build a query (‘All sports models made by Alpha Rodeo or Beamer’, for example). The client passes the query to the server as a PCatalogQuery. Inside the PCatalogQuery are three array attributes that identify the category names, engine sizes and

Designing the Business Services

make names that the user is interested in. The server returns an array of matching PCarModel objects. (The P stands for protocol – it helps the developers of the server objects to avoid any name clashes with heavyweight business objects. Arrays are used because they’re more compact than collections.) In order to reduce the amount of information returned to the client, the PCarModel objects do not contain any details, just the price, model number and universal identiﬁer. When the user asks for the details of a particular car model, rather than passing over the whole PCarModel, whose attributes the server must already have, the client passes over just the universal identiﬁer, as a parameter to the readCarModelDetails message.

10.5.3 Session Identiﬁers Often, in order to frustrate hackers, we need to restrict client access to privileged services. For example, for iCoot, we have privileged services, such as ‘Reserve a CarModel’, that require the client to have logged in and non-privileged services, such as ‘Read the index headings from the Catalog’, that are available to everyone, even hackers. Some client–server protocols, such as HTTP, have a standard challenge mechanism for displaying a log-on screen to the user: the user name and password typed in is validated on the server; if successful, the client is given a unique session identiﬁer. However, if we want to provide pluggable business services, i.e. services that will work with any kind of front end, we have to implement a mechanism ourselves. Figure 10.25 shows one way to do this. All the privileged services take a long number as their ﬁrst parameter: this must be a session identiﬁer created by the server layer, or use of the privileged service will fail. To get hold of a session identiﬁer, clients must use the logon message on the AuthenticationServer. (The steal parameter is used by the client to specify whether any existing session for this member should be terminated – this is part of the single-log-on mechanism.) When the logon method is called, the AuthenticationServer checks the membership number and password using the business layer. If the client’s credentials are correct, the AuthenticationServer generates a random number and associates it with the corresponding Member in the business layer. Subsequently, whenever a privileged service is invoked, the relevant server object can use the session identiﬁer to look up the Member before proceeding. Naturally, if the original credentials or the session identiﬁer are invalid, the client receives an error message. The session identiﬁers must be difﬁcult to fake – a randomly-generated 64-bit number will sufﬁce. Hackers generally do not bother trying to guess 64-bit random numbers because their chances of success are tiny. Another way of providing a portable privileged/non-privileged mechanism would be to make the session identiﬁer an object. Through encapsulation, this would give us more ﬂexibility in the kind of difﬁcult-to-fake information that we chose to employ.

309

310

Chapter 10

10.5.4 Business Service Realization Now that we have designed the business services (the messages on our server objects), we need to work out how they’re going to be implemented in terms of the business layer. To do this, we need to walk through the use cases, drawing sequence diagrams that show which messages need to be sent. For want of a better name, we’ll call this process business service realization. This is similar to the use case realization that we did during analysis, walking through the use cases and drawing communication diagrams to demonstrate that the business objects would support an implementation. We should use sequence diagrams to document our business service realization rather than communication diagrams, because sequence diagrams are more compact (we’re being speciﬁc about the implementation, so we have more information to show). Figure 10.26 shows a sequence diagram for the AuthenticationServer’s logoff method. Along the top are the objects involved in the interaction. Unlike objects in communication diagrams and object diagrams, the objects shown in sequence diagrams do not have underlined labels. Time ﬂows down the page, so, starting at the top, the Member actor is shown initiating the log-off via the AuthenticationServlet, which in turn sends the logoff message to the AuthenticationServer, which sends the ﬁndById message to the MemberHome, and so on. sd U12

Member

:Authentication Servlet

:Authentication Server

:Member Home

m:Member

:Internet Account

logoff() logoff(id) m=findBySessionId(id) id retrieved from browser

setSessionId(0)

session

setSessionId(0)

home page

Figure 10.26: Depicting business service realization

The dashed vertical lines on a sequence diagram are called lifelines (the lines don’t have to be dashed on a hand-drawn diagram). The vertical grey rectangles are called activation bars; they indicate when an object is executing a method. (The bars may be white or omitted

Designing the Business Services

altogether on a hand drawing). Each sequence diagram can be enclosed in a frame – a box with an operator in the top left corner, like the one we saw earlier for our design class diagram, but with a different operator. (All UML diagrams can be enclosed in a frame, the operator indicating the context for the diagram.) In the case of a sequence diagram, the operator is sd followed by the sequence diagram name. For a class diagram, we use the pkg operator, indicating the package that contains the classes. In this book, frames are only placed around the sequence diagrams and design-level class diagrams from iCoot. Frames can also be used to enclose loops and references to other sequence diagrams (see Appendix 3). Case Study iCoot sequence diagram A member of the iCoot team describes the interaction shown in Figure 10.26 as follows: When a Member actor elects to log off, their browser tells the AuthenticationServlet to logoff. The AuthenticationServlet then sends the logoff message to the AuthenticationServer, along with the Member’s id, which it retrieves from the browser session. The AuthenticationServer uses the id to retrieve the corresponding Member from the MemberHome. The AuthenticationServer then sends the setSessionId message to the Member, with 0 as parameter, which is passed on to the Member’s InternetAccount. The InternetAccount stores the session identiﬁer 0, to indicate that the Member has logged off. Finally, the Member actor is presented with the home page so that they can access customer services or log on again.

For business service realization, we need to show the messages that ﬂow between the server objects and the business layer objects, but we don’t need to show the inner workings of the business layer objects themselves. Depending on the size of the system, it may or may not be feasible to show every business service on a sequence diagram. In general, we should expect to show the most important scenarios – the use cases will help us to decide which ones are the most important. Within each use case, there may be several normal scenarios and several abnormal ones: for our purposes, it is enough to show the normal path for the most common scenario. As we draw sequence diagrams, we will discover more and more messages on the interfaces of our business layer objects. The resulting messages on the business layer objects are fed into the speciﬁcation phase, along with any others that we come up with using intuition,

311

312

Chapter 10

experience, libraries, patterns, frameworks and guesswork. In the speciﬁcation phase, we try to complete the object interfaces and describe the required behavior (see Chapter 12).

10.6 USING PATTERNS, FRAMEWORKS AND LIBRARIES A pattern (see Chapter 11) is a portable solution to a small programming problem – in the object-oriented sense, a pattern is a handful of collaborating objects. Patterns allow developers to work faster and produce better code by not duplicating effort. Some fundamental patterns have been mentioned at other points in this book: singletons (where a class has only one instance); factories (for creating objects); homes (for creating objects and ﬁnding existing ones); and states (representing an object’s life cycle). Each pattern has a name, a description and some examples of its use. Every developer should familiarize themselves with the common patterns. In some ways, a framework is similar to a pattern: it’s a way of putting together part of a system. Frameworks, however, have two important differences: ﬁrstly, they tend to be much larger; secondly, some of the code is already written for you (this may take the form of partially implemented classes or code-generation tools). Just as with patterns, you should seek out frameworks that have been designed for your particular problem domain, so that you can avoid writing unnecessary code. One of the most comprehensive and popular frameworks is Enterprise JavaBeans. The EJB framework allows developers to construct middle-tier business logic without having to write any code to deal with persistence, transactions, security, concurrency, distribution or thread safety. Although it’s a large framework, it’s certainly worth studying. A library is a collection of prewritten classes that can be used off-the-shelf. You should make yourself aware of the libraries available for your problem domain, so that you can avoid writing new classes when suitable ones already exist. The Java 2 Platform library is a good example. It comes in three varieties, Enterprise Edition, Standard Edition and Micro Edition, targeted, respectively, at large-scale systems (for example, e-commerce), mediumscale systems (for example, desktop publishing) and small-scale systems (for example, PDA e-mail clients).

10.7 TRANSACTIONS Every concurrent system, if it’s going to be robust, must be built around transactions. A transaction, also known as a unit of work, is used to fence off a number of database accesses.

Transactions

(A database access means ‘reading some data’ or ‘writing some data’.) Transactions are fairly complicated, so a detailed discussion is beyond the scope of this book. Transactions are used to guarantee that: • Information in the database doesn’t become corrupted by system problems. We want to make sure that the database moves from one consistent state to another: data is never partly updated; it’s either completely updated or not updated at all. • Clients don’t get hold of out-of-date information. We want to avoid situations such as the following: client A reads a customer’s address, client B modiﬁes the customer’s address, client A takes some action based on the old address. A database client starts a transaction, accesses some data and then commits the transaction. If the commit is successful, all updates made since the start of the transaction are ﬂushed to the database (while the transaction is active, the updates are merely pending) and the client can be sure that it acted on up-to-date information. If the commit fails, because of a system problem or a clash with another transaction, the client can roll back the transaction, discarding any pending updates (and, therefore, any data that was updated on the basis of out-of-date information will remain unchanged). Because the DBMS guarantees that the accesses within a transaction all succeed or all fail, we know that the database will move from one consistent state to another. Obviously, the DBMS must ensure that all accesses take place inside a transaction. Transactions can be short – wrapped around access to a single row – or long – wrapped around several accesses to related rows. A DBMS generally allows the client to choose between short and long transactions: by default, every access to the database is wrapped inside a new (short) transaction; alternatively, the client can opt to start, commit and rollback a (long) transaction manually. In theory, transactions can be nested: this allows us to wrap small units of work inside larger ones. In practice, however, most relational DBMSs do not support nested transactions.

10.7.1 Pessimistic and Optimistic Concurrency Concurrency control (using transactions to control simultaneous access to data from multiple clients) can be pessimistic or optimistic. With pessimistic concurrency, the DBMS guarantees that no other transactions can perform conﬂicting accesses while a transaction is active. With optimistic concurrency, transactions access at will but, when a transaction is committed, the DBMS checks that no other transactions have performed conﬂicting accesses in the meantime. By default, relational databases use pessimistic concurrency. Optimistic concurrency is more common in object-oriented databases, although it is sometimes provided as an option by object-oriented frameworks that sit on top of relational databases. (One way to implement optimistic concurrency for a relational database is to reread a row before updating it: if data

313

314

Chapter 10

in the row has changed, we know that another transaction modiﬁed it; however, this is only a partial solution.) To implement pessimistic concurrency, a relational DBMS locks data accessed within a transaction until the transaction has ﬁnished. For example, if a transaction modiﬁes the balance of account 123, the DBMS locks the row containing the new balance; other transactions are prevented from reading the row until the ﬁrst transaction ﬁnishes (and the lock is released). Locks apply to reads as well: for example, if a transaction reads the balance of account 456, other transactions won’t be allowed to modify the row containing the old balance until the ﬁrst transaction ﬁnishes (otherwise, the ﬁrst transaction might do half of its work relative to the old value and half of its work relative to the new value, which wouldn’t make sense). For performance reasons, relational DBMSs allow the client to relax the locking scheme, at the expense of some semantic accuracy.

10.7.2 General Guidelines for Using Transactions with Objects When mapping from an object-oriented model to a relational database, we have a conﬂict of interest. On the one hand we have a complex object graph, whole chunks of which we might want to process inside a single long transaction. On the other hand, we have a relational database with a pessimistic locking scheme by default and object data spread over many tables: this suggests that we should use short transactions, in order to avoid locking up large parts of the database. (This is why optimistic concurrency is popular with object-oriented databases and object-oriented frameworks.) Getting around this conﬂict requires skill and experience. If you’re lucky, you’ll be using a sophisticated framework such as EJB that solves most of the problems for you. If you’re performing the mapping by hand, consider the following advice: • Organize objects and access paths to reduce overlap. For example, with iCoot, we can make sure that each member is logged on only once. Thus, overlaps will only occur on the rare occasion that an assistant accesses a member’s data while that member happens to be logged on. • Use primary keys so that database accesses are well focussed. The trouble with most relational DBMSs is that they lock an entire table if they can’t be sure which rows are being accessed. For example, if ID is a primary key in the CUSTOMER table, a focussed query such as ‘Get me the name of customer 789’ will lock one row; a query involving non-primary columns, such as ‘Get me customers with the surname ‘‘Bloggs’’’ might lock the whole table. So, make sure that you choose a primary key for every entity and tell the database about it; then, base your database accesses on primary keys whenever possible.

Handling Multiple Activities

• Keep transactions short. Although you may want to control the start and end of a transaction, so that you can make several object accesses in one go, don’t overdo it.

10.7.3 Transactions in Upper Layers Transactions have a ripple effect: the fact that they’re in the database layer usually becomes obvious in the persistence layer; once they’re obvious in the persistence layer, they usually become obvious in the business layer; and, once they’re obvious in the business layer, developers of the server layer must understand them and how to use them properly. Generally, we can’t hide transactions until we get out to the client: to do this, the server layer must encapsulate transactions inside simpliﬁed requests (business services).

10.8 HANDLING MULTIPLE ACTIVITIES Normally, when we use a computer, we want to be able to do several things at once – write a letter, read e-mail, run a lengthy computation and browse the Web, for example. We may want a server to do thousands of things at once (handling simultaneous requests from multiple clients). To this end, most operating systems permit multitasking: each program runs as an independent process with its own protected area of code (program instructions) and data (program variables). Some programming languages allow us to execute multiple tasks within a single process. These tasks are usually referred to as threads of execution or just threads. Normally, each thread represents an activity within the program, running alongside other activities. In this section, we’ll examine the issues surrounding multi-threading and how we can make our code thread-safe (this turns out to be most important for the business layer).

10.8.1 Controlling Multiple Tasks Each process managed by the operating system can be idle (waiting for user input perhaps) or active (performing some computation). Because we usually have more processes than CPUs, the operating system must share the CPU time between the active processes: the operating system allows each process to run for a small amount of time and then moves on to the next. This time-slicing is controlled by a piece of software called a scheduler. We don’t need to know the details of the algorithm the scheduler uses to distribute the CPU’s time among processes – we can just assume that each process gets its fair share. As an extra facility, most operating systems allow us to assign a priority to each process, so that some processes get more of the available time than others. For example, we might give a high priority to the detection of mouse clicks and a lower priority to user applications.

315

316

Chapter 10

Although any decent operating system will prevent processes from accessing each other’s code and data, it’s the programmer’s job to make sure that access to external resources (such as ﬁles and databases) is managed sensibly. For example, when a word processor opens a ﬁle, it can lock the ﬁle to prevent other processes from editing it at the same time; concurrent access to a highly shared resource such as a database is usually controlled through a combination of transaction management and business rules. From the perspective of an individual user, multitasking allows us to have multiple applications open at once on our desktop. We can switch between the applications at will, doing one thing at a time, or set off several tasks at once, each of which appears to ﬁnish independently. Multitasking also has the advantage that, having set off a lengthy computation, we don’t have to wait until it ﬁnishes before we do something else: for example, while we’re waiting for an Internet search to complete, we can check the time or collect our messages. From the server perspective, as well as being able to serve many clients simultaneously, multitasking gives us better throughput (the server deals with clients more efﬁciently). For example, imagine that a script called search.pl is used to execute Internet searches over HTML/CGI. Some searches issued by clients will execute quickly, in milliseconds perhaps, while others will take several seconds. If a client starts a long search and then a simpler search comes in from another client, the second search can execute immediately, without waiting for the ﬁrst one to complete.

10.8.2 Controlling Multiple Threads Threads are different from processes in that they all share the same data area within their process. Therefore, as well as protecting external resources, the programmer has to protect internal data. (The code area inside each process is normally hidden from threads by the run-time system, so we don’t need to take any special steps to protect it.) In all other respects, threads are just mini-processes: they’re controlled by a scheduler and we can assign different priorities to them. From a client point of view, multi-threading has the following advantages: • The user can run many applications at the same time and do many things within a single application: for example, in a single e-mail process, we can edit a message, be notiﬁed when new mail arrives, view a real-time clock, and so on. • The user can interact with the user interface even if the application is busy. For example, imagine a database querying tool where the user types in a query and presses the Retrieve button; then, while the query is executing, the user notices that they have made a spelling mistake in the query. If the query tool has only one thread, the user can’t edit the query until the useless results have been returned and displayed. If, on the other hand, we

Handling Multiple Activities

arrange for the user interface and the database query to run in separate threads, the user can issue another search before the ﬁrst one has ﬁnished: the application can kill the incorrect thread immediately and the incorrect results are never displayed. • The user interface can be updated even when the application is busy. Consider a query tool, running as a single thread. If the user initiates a search and then, before the results are displayed, resizes the application window, what happens? Well, grabbing the corner of the window and moving it across the screen is performed by the operating system (the desktop), so the window boundary will move as expected. However, the inside of the window has to be painted by our application: since the application is busy, the inside of the window won’t be repainted until the query results come back. The user sees a rather amateurish user interface that repaints at unexpected times. If we use a separate thread for the query, the user can resize the window while they’re waiting and the repainting will happen immediately. From a server point of view, multi-threading is good because: • It allows us to serve many clients simultaneously without the overhead of multiple processes. Processes are much more expensive to set up, execute and tear down than threads. For example, a machine that crashes when you ask it to run 1000 processes simultaneously may be perfectly happy running four processes with 250 threads each. For certain kinds of networked application, this is critical: for example, servlets run in a single process with multiple threads but CGI scripts, by default, run in multiple processes; thus, if we want the beneﬁts of servlets, we have to have multi-threading. • It reduces latency (idle time) in the server. For example, if a middle tier machine accesses a database server using a single server thread, the middle tier machine is idle while the query is executed on the data tier machine. With multiple threads, the middle tier machine can be doing other work while the query is executing. • It reduces time-outs. With some protocols, a client request will fail automatically if the server doesn’t respond within a certain length of time (say, two minutes). If all client requests have to queue, waiting to be served by a single server thread, we will get more time-outs (each request is lengthened by the time it takes to serve the requests that were already in the queue). With multi-threading, short requests have a faster turn-around: time-outs will only happen for network problems, server overloading and overly-complex requests. Ideally, the programming language and its run-time system handle the messy details of multi-threading – scheduling, priorities, time-slicing, etc. This way, the programmer just has to write the code that will be run by the threads and start them up.

317

318

Chapter 10

10.8.3 Thread Safety Multi-threading causes problems, because threads can be interrupted before they’re ﬁnished (to allow other threads to run). For example, consider the following scenario: Two threads A and B are accessing an object O. Thread A starts to read one of O’s ﬁelds, F, using a getter method. When A has read half of the value, the scheduler interrupts it so that B can run for a while. B starts to modify F, via its setter. The scheduler allows B to ﬁnish its modiﬁcation before it wakes up A. When A wakes up, it reads the rest of F. Thread A has now read half of the old value and half of the new value, which is clearly nonsense. This kind of data corruption applies to external resources too (imagine if A were reading a text ﬁle and B were modifying it). When we access data in a database, the DBMS provides us with a sophisticated transaction mechanism to make sure that data isn’t corrupted. However, inside multi-threaded code, we have to protect the data ourselves. The key to protecting data in an object-oriented program is to make sure that each piece can only be accessed via a single object that manages the data. Then, as long as we ensure that only one thread accesses the object at a time (mutual exclusion), we know that the data will be safe. Preferably, our programming language will allow us to enforce mutual exclusion (we’ll see how it’s done in Java shortly). Code that is safe for multi-threaded use is said to be thread-safe or MT-safe (as opposed to ‘not thread-safe’ or MT-hot). Generally speaking, we would like to make all of our objects thread-safe and multi-thread all of our applications.

Immutability An immutable object is an object whose data can’t be changed. Here the term data means: • • • •

The values of the object’s ﬁelds. The values stored in external resources managed by the object (such as text in ﬁles). The values inside any objects pointed to by the object. ...

In other words, for an object to be truly immutable, it must be impossible to change the object’s own ﬁelds and any data that can be reached by the object, directly or indirectly, internally or externally.

Handling Multiple Activities

Immutable objects have the advantage that they’re always thread-safe – since there’s no data that can be changed, there’s no data that can be corrupted. They’re also more efﬁcient (they can be shared transparently and kept in read-only areas of memory). Some languages provide facilities for ensuring immutability – C++’s const keyword and Java’s ﬁnal keyword are common examples. These facilities, however, tend to be partial. A better approach is to enforce immutability by programming style (not providing setters, locking ﬁles, and so on). Although immutable objects are a nice idea, most objects need to be mutable: for example, a Customer that didn’t allow us to change its address attribute wouldn’t be much use. Thus, we have to know how to make mutable objects thread-safe.

Fixed Values Working out how to make objects thread-safe is a challenge. To make matters worse, we usually have to reason about whole families of objects and how they will be used together. One reason for this is deadlock. Deadlock refers to the situation where thread A is waiting for thread B to do something while thread B is waiting for thread A to do something: both threads end up waiting for ever. In order to avoid deadlock, we have to think about how our objects collaborate and how threads will wander through them. One simple trick we can use to help achieve thread safety is to look for ﬁxed values in our objects. A ﬁxed value is an immutable ﬁeld: for example, a Math object might have a Pi value inside it that never needs to be changed. Fixed values, being immutable, are automatically thread-safe, so scenarios like the ‘interrupted read’ that we saw earlier, are not a problem. Having decided which of an object’s ﬁelds are ﬁxed, we can divide our object into two halves: ﬁxed values, which don’t need special code to protect them, and changeable values, which do. Fixed values only have to be immutable after the object in question has been created. In other words, we can manipulate ﬁxed values inside an object’s constructors: as long as we don’t change the value after the constructor has ﬁnished, everything will be ﬁne. The reason for this is that only one thread can get inside a constructor: the thread that asks the run-time system to create the object; no other thread can get inside the object while it’s being constructed, because it doesn’t exist yet. (This assumes that the constructor doesn’t make the object available to other threads while it is executing.)

Synchronization in Java We can solve most multi-threading problems by encapsulating each shared resource inside a single object. It is then the object’s responsibility to make sure that only one thread is allowed in at a time. Preferably, the programming language should support this mutual exclusion. For example, in Java, a method can be marked as synchronized: the run-time system guarantees that only one thread at a time can be active inside any of an object’s synchronized

319

320

Chapter 10

methods. This is achieved by associating a lock with each object, under the control of a monitor. The ﬁrst thread to arrive at one of the object’s synchronized methods is allowed in by the monitor, but other threads are locked out of the synchronized methods until the ﬁrst thread departs. Java’s mutual exclusion does not apply to unsynchronized methods: threads are free to run in and out of them at any time. Figure 10.27 shows a snapshot of a Java object in use, with four threads trying to get inside. This object has MT-hot values, which need protecting, and ﬁxed values, which don’t. The three threads that we’ve called T1, T2 and T3 are currently active; T2 is suspended outside method M2 because T1 has already entered the object through another synchronized method (M1). For this scheme to work, the programmer must ensure that only the code inside M1 and M2 accesses the MT-hot values. The ﬁxed values, on the other hand, can be accessed from any method. Thus, in order to make a Java object thread-safe, we need to synchronize access to all the MT-hot values. In practice, this requires experience and hard thinking in order to avoid deadlock and unnecessary synchronization (this is important because mutual exclusion can reduce an object’s throughput).

Case Study Thread safety in iCoot So, how do we address the thread-safety of iCoot? We can consider each layer separately (another advantage of using layers): • In keeping with servlet programming style, our servlets (part of the distributed interface) are state-less, and therefore MT-safe. Session data (such as the PMember for the current user) is stored in one HttpSession per client and is protected using Java’s synchronization mechanism (using synchronized blocks). (As part of the standard HTML/CGI-plus-servlets mechanism, the Web server stores the session objects and the Web browsers store the session identiﬁers.) • Our pluggable server objects are also state-less and therefore thread-safe: each individual business service returns a response (as protocol objects) that is detached from the business layer and used only by the client that requested it. • The business layer has to be made MT-safe, by careful programming, so that multiple threads can run through it from the server layer without corrupting cached data read from the database layer. Incidentally, the database layer is concurrent-safe by default, courtesy of its transaction mechanism – the programmer simply has to make sure that a transaction is created at the start of each business service and committed at the end.

Further Reading

T2

T1

= Object M1

= MT-hot values

M2

= Fixed values = Synchronized method T4

T3

= Other method = Thread

Figure 10.27: Synchronization in Java

10.9 SUMMARY In this chapter, we looked at subsystem design – the process of deciding exactly what objects we are going to implement and what interfaces they should have: • We considered the design of the business layer and how to derive it from the analysis class model. • We saw how an object model could be mapped onto a relational database schema. For the sake of simplicity, we didn’t look in detail at how we would write the code to perform the actual mapping at run time. • After a brief look at tips for designing user interfaces, we discussed how to group the facilities offered by the middle tier into business service classes that hide the complexities of the business layer (for the beneﬁt of different kinds of user interface). • We considered the importance of looking for patterns, libraries and frameworks to avoid writing fresh code. • We looked at the issues surrounding database transactions and multi-threading (intra-process concurrency), the concepts involved and an example of mutual exclusion in Java.

FURTHER READING When considering how to map an analysis class model into design, it’s important to keep good theory and practice in mind. For a theoretical discussion by Bertrand Meyer (but which

321

322

Chapter 10

is readable nonetheless), see [Meyer 97]. Martin Fowler’s popular book [Fowler 03] explains some of the fancier parts of class diagram notation; more can be found in the UML Speciﬁcation [OMG 03a]. For a discussion of best practices for writing Java source code, see [Bloch 01]. Scott Ambler, an agile methodology enthusiast, provides comprehensive coverage of object-to-relational mapping in [Ambler 03]. In [Constantine and Lockwood 99], you will ﬁnd advice from Larry Constantine on how to design user interfaces according to the way the system is used (based on use cases, of course). J2EE covers all parts of multi-tier design and implementation, from GUIs and HTML front ends, through to servlets and EJBs on the middle tier and an object-to-relational mapping generated automatically by tools. Patterns for use with J2EE are described in [Alur et al. 03]. For a discussion of thread safety in Java, and some reusable patterns, see [Lea 99].

REVIEW QUESTIONS 1. What kind of diagram is shown in Figure 10.28? Choose only one option.

sd17

:Member Applet

:Authentication Server

:Member

:Member Home

logon() logon(“M1”, “xyz”, true) find(“M1”) 467:Member a:=isGoodMember()

b:=getPassword()

p:=PMember(792,467)

p p

Figure 10.28: Used with Review Question 1

p:PMember

Answers to Review Questions

(a) (b) (c) (d) (e) (f) (g)

State machine diagram. Activity diagram. Class diagram. Use case diagram. Sequence diagram. Communication diagram. Deployment diagram.

2. Currently, what is the most common type of database management system? Choose only one option. (a) (b) (c) (d) (e)

Network. Relational. Object-oriented. Hierarchical. Indexed ﬁle.

3. In UML diagrams, how are class messages distinguished from instance messages? Choose only one option. (a) (b) (c) (d)

Class messages are shown in brackets. Class messages are shown in italics. Class messages are underlined. Class messages are shown with the keyword static .

4. What is meant by the term ‘deadlock’? Choose only one option. (a) Two processes or threads refuse to talk to each other. (b) An object’s monitor allows its lock to terminate early. (c) An object is waiting for a resource, which is being used by an object waiting for a resource used by the ﬁrst object. 5. What is a ‘thread’? Choose only one option. (a) An independent process running on a node, with its own memory and IO. (b) An activity within a process that shares memory with other activities. (c) A designer’s thought process.

ANSWERS TO REVIEW QUESTIONS 1. The diagram in Figure 10.28 is e. Sequence diagram. 2. The most common type of database management system is b. Relational.

323

324

Chapter 10

3. In UML diagrams, class messages are distinguished from instance messages because c. Class messages are underlined. 4. The term ‘deadlock’ means c. An object is waiting for a resource, which is being used by an object waiting for a resource used by the ﬁrst object. 5. A ‘thread’ is b. An activity within a process that shares memory with other activities.

11 Reusable Design Patterns

W

hen writing software, as with any discipline, it’s important to avoid duplicating effort – if somebody has discovered a perfectly good solution to a problem, we want to exploit their knowledge and experience, rather than waste time producing another solution. Always remember, object-oriented programming is about not writing code: the more we reuse, the more skilled we are.

Learning Objectives Understand what a pattern is and what it means to a developer of object-oriented software.

•

Examine some of the more common patterns.

Identify patterns that apply to our own software and how to use them together.

Chapter Outline 11.1 Introduction 11.2 A Pattern Template 11.3 Common Design Patterns 11.4 Using Patterns 11.5 Discovering, Combining and Adapting Patterns 11.6 Summary Further Reading

11

328

Chapter 11

11.1 INTRODUCTION Design patterns are one way that developers can avoid duplicating effort: they allow us to apply the knowledge and experience of other developers to our particular problem. Patterns also allow us to communicate our own knowledge and experience to others. Each pattern is a description of a particular way of doing something that has proved effective in the real world. When design patterns were introduced to the software community, they created a lot of interest. As a result, they have now been applied to many other areas: • • • • • • • • • • •

Human–computer interaction Concurrency Reuse Teaching/learning object technology Distributed computing Project management Web sites Risk management Anti-patterns (things going wrong) Problem-solving Organizations

11.1.1 A Brief History of Patterns Originally, patterns had nothing to do with software. In the 1960s, Christopher Alexander, an architect of buildings, started writing about patterns in architecture and urban planning. Having studied human language, Alexander considered a sentence to be made up of reusable pieces or patterns that are employed by all of us to record knowledge, construct complex communications and solve problems. Because of its patterns, Alexander considered natural language to be a pattern language. Alexander believed that architecture and urban planning could be improved if a pattern language could be developed for them. Some of the improvements would result from having concise descriptions of the knowledge of experts, while more still would result from the improved communication, allowing members of the public to be involved in the process: it seemed reasonable that, to ensure a successful outcome, the eventual inhabitants of a space should inﬂuence its design [Alexander et al. 77]. In the late 1980s, when object-oriented programming was becoming more and more popular among researchers and practitioners, experts including Kent Beck, Ward Cunningham, Erich Gamma, Bruce Anderson and Richard Helm started to think about how patterns

Introduction

could be applied to software. As experienced designers (and Smalltalk programmers), these experts were well placed to start identifying and recording patterns of collaborating objects, ones that had shown themselves to be useful over time. The work of these enthusiasts eventually became centered around the Object-Oriented Programming, Systems, Languages and Applications (OOPSLA) conferences in the early 1990s. Ultimately, this effort led to a seminal book [Gamma et al. 95], which contained a general discussion of patterns, along with a catalog of 23 patterns ready for wider use – it’s a mark of the importance of this book that, a decade later, its 23 patterns are still considered to be fundamental. The authors are often referred to as the Gang of Four.

11.1.2 Software Patterns Today Today, as a result of the original book, the term ‘design pattern’ is mostly used to refer to software design patterns. Often, once the context is obvious, developers simply refer to patterns. Real enthusiasts still use the term pattern language to describe existing patterns and the way that they’re used to record knowledge, to communicate and to construct solutions. The original inventors of design patterns were true to Alexander’s vision, involving end users in the development process. However, patterns are not normally used in that way today. This may be because software design, even small pieces of software design, can’t easily be understood by non-programmers. Or, it may be because we don’t yet know enough about software to produce patterns that are simple enough for the lay-person. (Involving end users is still desirable, but we tend to use other artifacts such as use cases, analysis class diagrams and GUI sketches.) We’ve seen before how software is an inexact science, if a science at all. The gaps between the patterns are much larger than the gaps between any patterns that might exist in natural language or architecture. This may be because software design is too difﬁcult to condense into a set of interlocking patterns, or it may simply be that we don’t know how to do it yet. Patterns can certainly be used by software developers to record knowledge, so that other developers don’t have to duplicate their effort. Identifying and describing a pattern takes skill and experience; even experts should not try to describe a new pattern until they have demonstrated its effectiveness in several different applications. As well as recording knowledge, a pattern can be used as a pre-fabricated part of a new design. Because patterns are independent of programming language and application domain, they can’t be used simply as they are, in the same way that a framework or a library can. Patterns need to be ﬁne-tuned for each particular situation. Nevertheless, they still save us a lot of time. Patterns are also used widely to document a solution. For example, rather than trying to describe how ‘messages that are sent to this object are forwarded over the network to the real implementation object, which has a similar interface’, we can simply record that ‘this

329

330

Chapter 11

object is a network Proxy’. For a complicated pattern, the name of the pattern can be worth a thousand words.

11.2 A PATTERN TEMPLATE Since patterns are meant to be used widely, they need to be presented in a widely accepted format. Ideally, the format will be easy to read and understand; but, in the name of completeness and correctness, each pattern description may end up being rather formal. This is the case with [Gamma et al. 95]: although each pattern is essential and well described, the end result is more of a reference work than a tutorial. In this book, the pattern descriptions are informal, showing the purpose, the structure and the mechanism, with code examples and illustrations. [Gamma et al. 95] describes a template that can be ﬁlled in by pattern authors. In order to give you a ﬂavor of how full-blown patterns should be described, the template headings are reproduced below along with a few words of explanation for each: • Pattern Name: A short name for the pattern (usually one or two words), Memento for example. Obviously, this name should be indicative of the pattern’s purpose. It should also be unique within the application area. • Classiﬁcation: Each pattern is classed as creational (concerned with how objects are created), structural (concerned with putting objects together into a larger structure), or behavioral (concerned with collaborations between objects to achieve a particular goal). • Intent: A short description (one or two sentences) summing up what the pattern is for, for example, ‘Without violating encapsulation, capture and externalize an object’s internal state so that the object can be restored to this state later’. • Also Known As: Aliases for the pattern. • Motivation: A description of a design problem that is solved by use of the pattern. • Applicability: Areas where this pattern can be applied and how to recognize those areas. • Structure: One or more class diagrams and sequence diagrams that illustrate how the pattern works. [Gamma et al. 95] uses OMT (a predecessor of UML); obviously, UML will be used here. • Participants: Short descriptions of the objects involved and what each one does (what its responsibilities are). • Collaborations: Description of the collaborations between the participants. • Consequences: Beneﬁts and shortcomings of the pattern (plus advice on what to do about the shortcomings). • Implementation: Advice on implementing the pattern, including useful tricks and things to avoid.

Common Design Patterns

• Sample Code: Full implementations of the pattern. In [Gamma et al. 95], the implementations are in C++ or Smalltalk – Java is used in this book. • Known Uses: Where the pattern has been applied in the real world. • Related Patterns: The patterns that are similar to this one and exactly how they differ. Also, which patterns can be valuable when used with this one.

11.3 COMMON DESIGN PATTERNS As a general principle, we must be familiar with the core patterns before trying to implement anything other than a trivial program. This section gives informal descriptions of the most common patterns. Complete, formal descriptions are available elsewhere but the information here is enough to provide a good grounding. You are encouraged to investigate other patterns for yourself, at least the other well-known ones (Prototype, Bridge, Builder, Memento, Command, Decorator, Chain of Responsibility, Interpreter, Mediator and Visitor).

11.3.1 Observer Deﬁne a one to many dependency between objects so that when one object changes state, all its dependents are notiﬁed automatically. [Gamma et al. 95] Often, the state of one object depends in some way on the state of another – a common case is a GUI component displaying the state of an entity. For example, Figure 11.1 shows a tool for previewing a car that a customer is considering buying. As each of the car’s attributes are varied by the customer, the picture of the car and the displayed price vary accordingly. The GUI object itself, which is an instance of class CarView, is composed of three panels, each displaying some information about the car. Under the covers, the GUI is supported by a Car object which has a ﬁeld for each of the possible options (with corresponding getters and setters). The price is derived from the other attributes and made available with getPrice. Where should we put the logic to detect when the car display needs updating? If we put it in the GUI, the GUI programmer must have knowledge of the way a Car operates – the sunroof affects the price but the color doesn’t, for example. This approach would make the GUI programmer’s life more difﬁcult. It would also spread knowledge around the system, making it harder to manage. If we put the knowledge in the entity instead, how can we make sure that the GUI is updated at the appropriate time? We could make the car entity aware of the GUI; however, we want the entity to be developed independently of the GUI so that it can be reused in many car applications – coupling it to a particular GUI would limit its usefulness.

331

332

Chapter 11

aCar

Color

Magenta

Doors

5

Sunroof

No

CD Player

Yes

-color = magenta -noOfDoors = 5

1

-sunroof = false -cdPlayer = true

Price $36,111

aCarView

Figure 11.1: A GUI for car buyers

So we seem to have a paradox: we want the knowledge about changing attributes to be in the entity, where it belongs; however, we can’t put it in the entity, because then the entity would be coupled to the GUI. We can solve this paradox with the Observer pattern. The basic idea behind Observer is to give all GUIs a simple interface that allows an entity to signal a change of state: this means that there is only a light coupling between the entity and the GUI. (The analogy is that the entity goes about its business, more or less oblivious to any object that happens to be watching what it’s doing.) Let’s look at Observer in detail, an abstract deﬁnition ﬁrst followed by an implementation of the car example. In Figure 11.2, the object with changing attributes is referred to as the subject, while an object that observes the changes is called an observer. The Observer class has a simple interface consisting of one update message – this message, which will be sent by the subject when an attribute has changed, provides an opportunity for an observer to refresh itself. The subject, for its part, maintains a collection of Observer objects that have registered an interest in changes – registration is achieved using the addObserver message. (If necessary, observers can be unregistered at a later date using the removeObserver message.) The subject also has a protected notify message that iterates over the observers, sending the update message to each one (for the sake of predictability, we can ensure that the update messages arrive in the same order that the observers were registered). The Subject class, despite having a complete implementation, is designed to act as a superclass – the reusable implementation is not much use without some attributes to

Common Design Patterns

Send update() to every registered observer

Subject

<> Observer

-observers +addObserver(:Observer) +removeObserver(:Observer) #notify()

1

*

+update()

ConcreteSubject

ConcreteObserver

-attribute:X +update() +getAttribute():X +setAttribute(:X)

Do something with subject's attribute

Set attribute and call notify()

Figure 11.2: Observer class diagram

modify – so we mark the class as abstract. Anyone wishing to implement an observable object can make their class inherit from Subject, as has been done here with ConcreteSubject. The implementor of ConcreteSubject must ensure that notify is called whenever attribute values are changed – inside setAttribute for example – so that the update message is sent to every observer. (In order to avoid unnecessary notiﬁcations, the implementor must make sure that this only happens when the new attribute value is different to the old value – most of the time, because we’re dealing with attributes, different means not equal, rather than not identical.) The Observer class has no concrete methods (because we don’t know what an Observer will need to do to refresh itself). Thus, Observer can be an interface (a pure abstract class). When we write an Observer class, it needs to inherit from Observer and implement the update

333

334

Chapter 11

anObject

aConcrete Observer1

aConcreteSubject

aConcrete Observer2

addObserver(aConcreteObserver1) addObserver(aConcreteObserver2)

setAttribute(anX)

notify()

update()

getAttribute()

anX update()

getAttribute()

anX

Figure 11.3: Observer sequence diagram

method to refresh its contents in some way – this is illustrated by the class ConcreteObserver in Figure 11.2. Typically, a ConcreteObserver will use ConcreteSubject getters to retrieve new attribute values during the execution of the update method. The sequence of messages for the Observer pattern is illustrated in Figure 11.3. Initially, anObject registers two observers with aConcreteSubject. Some time later, anObject modiﬁes aConcreteSubject’s state using setAttribute. Having recorded the new value internally, aConcreteSubject sends itself the notify message. (A stacked activation bar has been used, in ofﬁcial UML style, to show the duration of the notify method.) Inside the notify method, the update message is sent to aConcreteObserver1 and then to aConcreteObserver2. Inside the update methods, each observer uses getAttribute to discover the current state of the subject. The ConcreteSubject knows that it’s dealing with Observer objects, but not ConcreteObserver objects. (The parameter to each of addObserver and removeObserver is an Observer.) Thus, the

Common Design Patterns

ConcreteSubject is

coupled to the Observer class but not to ConcreteObserver: the two abstract classes provide the glue that makes the subject independent of the observer. Obviously, the observer is tightly coupled to the subject, but since the observer is usually in the layer above (in the subsystem design sense), that’s not a problem. The abstract classes, since they’re generic, can be added to a class library. Applying Observer to the car display tool in Figure 11.1 gives us the class diagram shown in Figure 11.4. A single observer, aCarView, registers itself with the subject and also acts as the source of updates. The message ﬂow is illustrated in Figure 11.5. Initially, the addObserver message is used for registration. Then, some time later, the user elects to add a sunroof: this results in the setSunroof(true) message being sent to aCar. Assuming that the sunroof attribute had been false, the update message is sent to aCarView inside notify. Inside the update method on aCarView we read the current attribute values using the getters.

Subject

<> Observer -observers

+addObserver(:Observer) +removeObserver(:Observer) #notify()

1

*

+update()

Car

CarView

-color:Color "... +update() +getColor():Color +setColor(c:Color) +... +getPrice():Money Set the color variable to c, then call notify()

Refresh the display from the current attribute values

Figure 11.4: Car observer class diagram

Although Observer is often used to allow a GUI to update itself when an entity changes, we can use the pattern elsewhere too, especially in a layered system. For example, entities

335

336

Chapter 11

aCarView

aCar addObserver(aCarView)

setSunroof(true) notify() update() getColor() magenta getSunroof() true getPrice() €38550 getNumberOfDoors() 5 getCDPlayer() true

Figure 11.5: Car observer sequence diagram

in the business layer may be observed by controllers in the control layer (controllers act as translators between the business layer and the user interface layer). In turn, the controllers can be observed by GUI components. (See Chapter 8 for an alternative way of signaling between layers, using events, that has the same effect as Observer.) In a multi-tier system, with a real GUI (as opposed to an HTML-based interface), Observer is invaluable on the client side. However, on the server side, we often have to deal with

Common Design Patterns

multiple threads and the extra complexity makes it difﬁcult to employ Observer. In theory, we could also use Observer to signal updates from the server to the client; however, in this case, the problems of multiple threads are exacerbated by multiple machines and network issues (such as time-outs and crashed clients). In short, it’s best to restrict our use of Observer to the client side, and then only to use it as a means of passing notiﬁcations from one layer to the layer above. In all other cases, we should avoid Observer, for the sake of simplicity and safety. (Exceptional cases on the server can be dealt with by other means, such as callbacks from the operating-system clock and machine-to-machine messaging, as exempliﬁed by the Java Messaging Service.)

11.3.2 Singleton Ensure a class has only one instance, and provide a global point of access to it. [Gamma et al. 95] From time to time, it’s useful to have an object that is the only instance of its class. Often, this will be because the object represents a unique component in the system or in the application domain – for example, in a system that stores data in a database, the database itself can be represented as a single instance of a class DB. Another common case is where the singular instance needs to be shared between separate parts of the system, saving memory space and creation time – for example, we might use a single Calendar to answer all questions such as ‘How many days are there in February?’, rather than creating lots of Calendar objects. In the ﬁeld of design patterns, an object that is the sole instance of its class is called a Singleton. In order for a singleton to be useful, it should satisfy the following three criteria: • It must be easy to ﬁnd. • It must be impossible for anyone to create another one. • It should not be created until it’s needed. In order to make the singleton easy to ﬁnd, we can store it in a class ﬁeld, accessed via a class message. This is reasonable since, in object-oriented programming, classes have long been used as the entry point for shared data and services. To ensure that only one singleton can be created, we can make sure that the code used to create the singleton is inaccessible to ordinary code by making it private. It is less important that the singleton is not created until it is needed but it is still desirable, in order to avoid creating and initializating an object that is never used. Client programmers won’t issue instructions to ‘create the object now’, so we can choose

337

338

Chapter 11

between creating the object when the system starts or creating it on-demand – typically, we choose the latter and use a simple technique called lazy initialization where conditional logic is used to check, at the point of access, whether the singleton has yet been created. Figure 11.6 shows a Singleton implementation as a class diagram. Here, we have a private class variable called instance which points to the Singleton object – this object is created on-demand by the getInstance class method (remember that UML distinguishes class elements from object elements by underlining them). We also have a constructor, indicated with the <> keyword. Since the constructor is private, it can’t be used anywhere outside the class. Finally, the Singleton class includes deﬁnitions of the instance methods, instanceMethod1, instanceMethod2 and so on. (Of course, our Singleton could also have ﬁelds, just like any other object.) In Figure 11.7, you can see a less abstract example of a singleton in the form of a calendar. Despite the new class name, the naming convention getInstance has been retained. -instance

1 Singleton 1 -Singleton() <> +getInstance():Singleton +instanceMethod1() +instanceMethod2() ...

Create instance, if necessary, and return it

Figure 11.6: Singleton class diagram

You may be feeling, at this point, that the class diagrams you’ve just seen are difﬁcult to understand, because they mix instance and class concepts in the same box. This is a sad fact of life for class diagrams in general, but one which it is worth getting used to. However, for the sake of clarifying the Singleton pattern, we can take an alternative view of what’s going on (see Figure 11.8), which you may ﬁnd easier to appreciate. Here, we consider the class to be a separate object in its own right, hosting the constructor (Calendar), the class ﬁeld (instance) and the class method (getInstance). Meanwhile, the singleton object hosts instance ﬁelds and instance methods. (Some programming languages, such as Smalltalk and

Common Design Patterns

-instance

1 Calendar 1 -Calendar() <> +getInstance():Calendar +getFirstDay():Day +getDayLength(:Day):Time ...

Create instance, if necessary, and return it

Figure 11.7: Calendar class diagram

Calendar class 1 -Calendar() +getInstance() 1

-instance

Calendar Create instance, if necessary, and return it

+getFirstDay(): Day +getDayLength(:Day): Time ...

Figure 11.8: Calendar as a class object

Java, do indeed model each class as a distinct object.) In UML terms, the class that has been labeled Calendar class is a metaclass: rather than using the ofﬁcial UML keyword, the naming rules have been bent a little, to make things clearer. Figure 11.8 gives rise to the object diagram in Figure 11.9. Here, creation of the singleton and access to it is managed by the class; the singleton itself is like any other object: it has a name, a type and, usually, some ﬁelds.

339

340

Chapter 11

Calendar class

-instance

aCalendar

Figure 11.9: Calendar object diagram

In programming terms, the implementation of the Singleton pattern is easy. In order for programmers to make use of a singleton, they must ﬁrst identify the class and then invoke (getInstance). The way this is done varies from language to language. Here it is in Java: public class Calendar { private static Calendar instance; // static means "class field" private Calendar() { // Must declare a private constructor // Initialize any fields here } public static getInstance() { if (instance == null) { instance = new Calendar(); } return instance; } // Declare any instance fields here... // And now for the instance methods: public Day getFirstDay() { ... } public Time getDayLength(Day aDay) { ... } ... }

The Java code fragment below shows the calendar being used to discover the ﬁrst day of the year, so that we can print it out for the user: Date d = Calendar.getInstance().getFirstDay(); System.out.println("The first day of the year is: " + d);

In Chapter 2, class ﬁelds and class messages were identiﬁed as a way of managing centralized data and services. As you can see from the calendar example, Singleton also provides access to centralized data and services. So which is better? Generally speaking, Singleton is the better choice, for two reasons: • Class elements are not very object-oriented because most programming languages do not support inheritance and redeﬁnition for them. This makes it difﬁcult to craft a

Common Design Patterns

hierarchy of different kinds of Calendar, such as Calendar, GregorianCalendar, JulianCalendar, IslamicCalendar, and so on. • As designers and programmers, it’s quite enough for us to have to deal with objects at run time, why should we have to deal with classes too? (Okay, so Singleton does use one class ﬁeld and one class method, but that’s as far as it goes.)

If you ﬁnd yourself wanting to use class elements, you should say to yourself ‘Perhaps what I really need is a singleton’.

11.3.3 Multiton Singleton is undoubtedly useful and you should expect to encounter it often. However, there is a related pattern, dubbed Multiton, that doesn’t appear in [Gamma et al. 95] (a form of it does, however, appear at www.patterndigest.com). The term multiton is a pun on ‘multi-valued singleton’ although that is, admittedly, an oxymoron. If you prefer, a multiton is any type with a restricted set of values. (This is similar to an enumeration in languages like C or in UML, except that each value of such an enumeration is a primitive rather than an object, which makes them less useful.) For example, we are designing a sales system for a car showroom that has cars available in the following ﬁve colors: sunset red, midnight blue, morning orange, noon yellow and afternoon grey. In object-oriented terms, we would like each color to be represented as an independent object with its own data and behavior (for example, the amount that this color adds to the cost of the car). We would like the ﬁve objects to be easily accessible. Furthermore, in order to avoid mistakes, we would like to stop programmers creating their own colors. This design begins to sound rather like Singleton but with every occurrence of ‘one’ replaced with ‘ﬁve’. Figure 11.10 shows a class called CarColor containing a lazily-initialized Map of CarColor objects, each retrievable by name. (A Map is a collection that allows us to insert and retrieve objects by name.) Our getInstance method now takes a String parameter allowing us to specify which of the ﬁve instances we’re interested in. CarColor also has a retrieveAllInstances message that we can use to get a list of all the instances – this would be useful for displaying the available colors in a GUI. Figure 11.11 shows CarColor separated from its contents, so we can see clearly what is going on: the class has a reference to the Map which, in turn, has references to ﬁve CarColor objects. Because the constructor is private, we know that client programmers can’t create their own CarColor objects. Multitons are useful in languages that don’t provide a special syntax for declaring types with a restricted set of values. Java has a form of multiton within the syntax of the language itself, so you don’t have to write your own.

341

342

Chapter 11

CarColor -colorMap:Map -name:String -cost:Money -description:String -CarColor(name:String,cost:Money,description:String) <> +retrieveAlllnstances(name:String) +getAlllnstances():List +getCost():Money +getDescription():String

Return a list containing all the values from the map

Return the corresponding value from the (lazily-initialized) map

Figure 11.10: Car colors using Multiton

11.3.4 Iterator Provide a way to access the elements of an aggregate object sequentially without exposing its underlying representation. [Gamma et al. 95] Often, we need to be able to do something for every object in a collection. For example, if we have a collection of items in a warehouse, we might want to add the value of all the items together to calculate the total value of the warehouse contents – in other words, we would need to ‘add this value to the total’ for every item in the warehouse. We don’t normally have direct support for this kind of task in our programming language, so we must use objects and messages instead. Iterator is a simple pattern that allows us to retrieve items one at a time from any kind of collection, using a standard loop – once we have a reference to an object inside the loop, we can perform the desired operation. For unordered collections, we don’t care about the order in which the objects are retrieved, as long as each is retrieved exactly once; for

Common Design Patterns

blueCarColor

redCarColor name = "red" cost = €105 description = "sunset Red"

name = "blue" cost = €95 description = "Midnight Blue"

orangeCarColor aMap

CarColor class

greyCarColor name = "grey" cost = €85 description = "Afternoon Grey"

name = "orange" cost = €115 description = "Morning Orange"

yellowCarColor name = "yellow" cost = €100 description = "Noon Yellow"

Figure 11.11: Multiton object diagram

ordered collections, we would expect the objects to be retrieved in order. Iterator requires the collaboration of collection-class implementors, which must provide at least a method to create and return an iterator. Figure 11.12 shows the Iterator pattern, represented by the Iterator interface (since we don’t expect to be able to provide a reusable implementation at this level, the class can be an interface). Alongside Iterator is a basic hierarchy of collections: the top level class, Collection, has an abstract method called createIterator that creates and returns an iterator of the appropriate type; underneath Collection, we have an unordered variety called Bag and an ordered variety called List. We expect to have to provide more than one concrete implementation of Iterator, because it’s unlikely that we could write a single iterator that would work for every kind of collection (although some iterators may work for more than one). Let’s assume that each concrete collection has its own concrete iterator, hence the classes BagIterator and ListIterator in Figure 11.12. We have to rely on the cooperation of collection-class implementors, who must ﬁnd or implement an iterator that will work for their collection and return it from createIterator. Given the class model in Figure 11.12, and appropriate method

343

344

Chapter 11

Collection +createIterator():Iterator +add(:Object) +remove(:Object) +contains(:Object) ...

List

Bag

+createIterator():Iterator +add(:Object) +remove(:Object) +contains(:Object) +getItemAt(:int):Object ...

+createIterator():Iterator +add(:Object) +remove(:Object) +contains(:Object) +take():Object ...

-list

1

1

*

-bag

*

ListIterator

BagIterator

+first() +next() +hasMore():boolean +getCurrent():Object

+first() +next() +hasMore():boolean +getCurrent():Object

<> Iterator +first() +next() +hasMore():boolean +getCurrent():Object

Figure 11.12: Iterator class diagram

Common Design Patterns

implementations, the client programmer can write uniform iterating code, as in the following Java example: Collection items = new Bag(); // "new List()" would work just as well items.add(new Widget()); items.add(new FDoobrie()); Iterator i = items.createIterator(); // Iterator now at first element while (i.hasMore()) { Object item = i.getCurrent(); ... // Do something with item i.next(); }

It turns out that the implementation of ListIterator is straightforward. Here’s how it looks in Java: public class ListIterator implements Iterator { private List list; private int index; // Automatically set to 0 protected ListIterator(List l) { // Constructor list = l; } public void first() { index = 0; } public void next() { index = index + 1; } public boolean hasMore() { return index < list.getSize(); } public Object getCurrent() { return list.getItemAt(index); } }

To complete the picture for ListIterator, we need to persuade the implementor of List to add the following method to their class: public Iterator createIterator() { return new ListIterator(this); // "this" means "the current object" }

345

346

Chapter 11

It is common practice to provide some kind of two-way iterator for collections that keep their elements in order, so that clients can iterate in reverse order if they want to. We can achieve this with a sub-interface of Iterator, called TwoWayIterator, that adds messages called last and previous. Then, given concrete implementations such as ListTwoWayIterator, each ordered collection can provide a factory method called createTwoWayIterator, to complement createIterator.

11.3.5 Factory Method and Abstract Factory Deﬁne an interface for creating an object, but let subclasses decide which class to instantiate. Factory method lets a class defer instantiation to subclasses. [Gamma et al. 95] Factory Method should be mentioned at this point because of its importance. A new description isn’t needed, because we’ve already seen this pattern being used in the implementation of Iterator (with createIterator and createTwoWayIterator). At its simplest, a factory method creates and returns an object of some type. Factory methods are convenient for client programmers because: • The client doesn’t need to know the concrete type of the object being created. Instead, they work with a higher-level abstraction (always a good idea with object-oriented programming). In our example, client programmers would need to know about the Iterator and TwoWayIterator interfaces but not the concrete types ListIterator, BagIterator and ListTwoWayIterator. • The type can be changed from time to time, or from platform to platform, without affecting existing client code as clients don’t know the actual type of object being created (another example of loose coupling). • Clients don’t need to be concerned with the creation details of concrete classes (ﬁddly constructor parameters, for example). Factory Method ‘lets a class defer instantiation to subclasses’. This allows the implementor of Collection to work with a mechanism that hasn’t yet been deﬁned. In other words, Collection can provide an abstract createIterator method which has separate deﬁnitions in List and Bag. In a related pattern, called Abstract Factory, a class contains multiple factory methods – this is used to provide a single point of creation for whole families of objects. (The Abstract part indicates that we can build a hierarchy of factories.)

Common Design Patterns

11.3.6 State Allow an object to alter its behavior when its internal state changes. The object will appear to change its class. [Gamma et al. 95] As we saw in Chapter 7, objects sometimes have a complicated life cycle, one that is complicated enough for us to want to use a state machine to model their behavior. When we come to implement such an object, it’s a good idea to avoid coding all the details of the corresponding state machine inside the object’s own methods; otherwise the complexity would be spread throughout the code, where it would be difﬁcult to manage. The State pattern is a convenient way of implementing a state machine separately from the original object, allowing us to see and modify the state-related behavior easily.

Ready

payment made

Paid

choice made drink dispensed

Chosen

Figure 11.13: State machine diagram for a vending machine

Consider the state machine diagram in Figure 11.13, showing the activities of a typical vending machine that dispenses one of a number of cans of drink, all at the same price. When the machine is powered up and ready to go, it is in the Ready state, waiting for the price of a drink to be inserted by the customer. A customer walks up to the machine and puts in some money; when the money reaches the price of a drink (the payment made event), the vending buttons light up – the machine is now in the Paid state. Next, the customer presses one of the buttons (the choice made event) and the machine is in the Chosen state while it retrieves the drink. Finally, when the drink has been dropped into the collection tray (the

347

348

Chapter 11

drink dispensed event), the machine is ready for another payment. So, how can we turn this state machine into objects, using the State pattern? The general form of the State pattern is shown in Figure 11.14. Here, the object with the interesting life cycle is called Context. There are two events in Context’s state machine that have been translated into messages on Context – someEvent and anotherEvent. To avoid coding the effect of these events inside Context, a separate object (called ContextState) has been added to do the work: any message that is sent to Context will be passed on to ContextState. ContextState is just a superclass, however: the real activity is added in subclasses, one for each state that appears in the state machine diagram (StateA and StateB in this case). The basic idea is that, when Context is in state A, its state variable will be an instance of StateA and when it is in state B, its state will be an instance of StateB. Since all state messages sent to Context are delegated to the state object, programmers can put state-related activity in the subclasses of ContextState, where it’s easier to ﬁnd.

Delegated to the state object

Context ContextState +someEvent() +anotherEvent() #setState(:ContextState) ...

Other methods, not state-related

1

* -context

-state

#someEvent() #anotherEvent()

StateA

StateB

#StateA(:Context) <> #someEvent()

#StateB(:Context) <> #anotherEvent()

Do the activity

Figure 11.14: State class diagram

When a Context is created, it sets its state variable to an instance of StateA or StateB, depending on which is the start state. During this initialization, the state object’s context variable will also be set. For example, for a Context class written in Java, we could have:

Common Design Patterns

public Context() { // Constructor new StateA(this); }

And, for a ContextState, we could have: public ContextState(Context c) { // Constructor context = c; c.setState(this); }

Now, for each event that happens to cause a transition to another state, the implementor of the event method can use Context’s setState message to get the context into the new state, as in: public void anEvent() { ... // State activity, followed by: new StateB(context); }

Although the scheme described above means that we have two links, one from Context to ContextState and another from ContextState to Context, it does keep the state-transition behavior inside the state objects, where it belongs. All methods and constructors on the state classes are protected, because they should not be accessible to clients. The ContextState class is abstract, because only instances of its subclasses should be created. Now look at Figure 11.15, showing the vending example implemented using State (the constructors and setState have been omitted, for simplicity.) Anyone wanting to work with a vending machine simply needs to create a VendingMachine object and send it the pay, choose and dispense messages at the appropriate times. Internally, the VendingMachine starts with an instance of Ready, which is replaced with an instance of Paid when the pay message arrives, then with an instance of Chosen when the choose message arrives, and then reverts to Ready when the dispense message arrives. This entire process, including creation of the objects, is illustrated in Figure 11.16. Inside each of the state methods, on the concrete subclasses, we can implement any behavior we like. This could take the form of manipulations of the states themselves, or manipulations of the VendingMachine context (since we have a reference to the context, we can easily send it messages). You may think that the methods on a class such as VendingState would be abstract. However, if we make them abstract, implementors of the concrete subclasses have to redeﬁne methods that are not relevant to them (because each event can only occur in certain states). Thus, in our vending example, the writer of Ready would have to provide an implementation for choose and dispense, even though these events could never happen (because the drinkselection buttons are inactive until enough money has been inserted and because the machine won’t dispense a drink until a choice has been made). Although some sort of implementation

349

350

Chapter 11

VendingMachine VendingState +pay() +choose() +dispense() ...

1

* -vendor

-state

#pay() #choose() #dispense()

Ready

Chosen

#pay()

#dispense()

Paid #choose()

Figure 11.15: Vending machine class diagram

could be provided for the irrelevant methods, reporting an error for debugging purposes perhaps, the end result would be inconvenient and clumsy. A better approach is to provide an implementation for every method on VendingState that reports an error – this way, writers of subclasses only have to override the methods that can occur in that state. Figure 11.16 contains some new UML notation which deserves explanation. A lifeline shows how long an object survives in the sequence. We can show an object (such as Ready) at the point where it is created. If the object is created by another object, we can indicate that with a message sent from the creator’s activation bar to the new object’s perimeter (optionally labeled with the constructor details). It is also possible to indicate the end of an object’s useful life by putting a large black X at the end of its lifeline, perhaps with an incoming message such as close. This marker, a stop in UML, indicates that the object can be deleted or that it must not be used beyond this point. Another issue that often has to be addressed with state machines is state data, i.e. attributes that are relevant to particular states. The amount of money that has been paid and the drink that has been selected are examples from our vending machine. State data also ﬁts neatly into the State pattern: all we have to do is add the getters and setters to the Context and ContextState classes, just as we did with state methods. The ﬁelds and concrete getters

dispense()

choose()

pay()

VendingMachine()

client

dispense()

choose()

pay()

setState(b)

Ready(a)

:Ready

setState(c)

Paid(a)

:Paid

setState(d)

Chosen(a)

:Chosen

Figure 11.16: Vending machine sequence diagram

:Vending Machine

setState(e)

Ready(a)

:Ready

Common Design Patterns 351

352

Chapter 11

VendingMachine

VendingState 1

1

+pay() +choose(:Drink) +dispense() +getDrink():Drink

-vendor

-state

#pay() #choose(:Drink) #dispense() #getDrink():Drink

Chosen

Ready

-drink: Drink

#pay()

#dispense() #getDrink():Drink Paid #choose(:Drink)

Figure 11.17: Vending machine with state data

and setters can then be added to the subclasses that actually use the data. For example, Figure 11.17 shows the effect of adding the customer’s choice of drink to our vending machine (again, constructors and state-setting methods have been omitted, for the sake of simplicity). Here, the choice made by the customer is passed in as a parameter to choose and then passed on to Chosen via its constructor. The Java implementation of choose for the Paid class would be something like: public void choose(Drink d) { new Chosen(vendor, d); }

11.3.7 Facade Provide a uniﬁed interface to a set of interfaces in a subsystem. Facade deﬁnes a higher-level interface that makes the subsystem easier to use. [Gamma et al. 95] When implementing a subsystem or layer using objects, the number of objects involved can be relatively large. For example, for a system like iCoot, the business layer runs to dozens

Common Design Patterns

of classes, both new ones and ones that we’ve re-used from elsewhere. This is reasonable, because we try to break down the complexity of a sizeable programming task into manageable objects that collaborate as necessary. The alternative would be fewer, larger, objects: such large objects would be more difﬁcult to implement correctly and many would end up being multi-purpose (i.e. they would have weak cohesion). However, having produced lots of sophisticated objects with many connections, the clients of our layer or subsystem have a problem: how do they use the rich set of objects to perform a simple task without intimate knowledge of the internal interfaces and collaborations? The Facade pattern comes to the rescue here: each facade translates the complexity of some part of a subsystem or layer into a single object with a subset of the available services. In our iCoot system, for example, half a dozen server objects were introduced to translate the multi-threaded, transaction-oriented business objects into a much simpler ‘request–response’ protocol that can be used by all manner of networked clients. In the request–response protocol, when a simple request comes in from a client (Web browser or applet), the relevant server object uses whatever combination of business objects and messages is necessary to generate a simple reply. This way, the task of the programmers of the user interface is made easier, with the added bonus that the facades can be used in different kinds of interfaces with minimal extra coding. Figure 11.18 shows two facades providing a limited number of services for clients of a complex subsystem, alongside the more complicated alternative where clients use the subsystem objects directly.

:Facade1

:Facade2

Figure 11.18: Subsystems with and without the Facade pattern

353

354

Chapter 11

11.3.8 Adapter Convert the interface of a class into another interface clients expect. Adapter lets classes work together that couldn’t otherwise because of incompatible interfaces. [Gamma et al. 95] Adapter wraps one object inside another so that the ﬁrst object can be used in a different context. Often, we need to do this in order to connect objects that weren’t designed to be used together. The basic form of Adapter is shown in Figure 11.19. Here, the client and the adaptee were designed separately, but, in the name of reuse, we would like them to work together. To do this, we interpose an adapter that has the interface expected by the client – inside the adapter, we forward incoming messages to the actual messages on the adaptee, translating message names, parameters and return types as necessary. Inside each adapter method, we can also provide any behavior that the adaptee lacks or modify the existing behavior.

message1(p1) client

message2(p2) adapter

adaptee

Figure 11.19: Adapter object diagram

In Figure 11.20, we have a simple application of Adapter. Here, the design requires a class called Queue that holds a queue of objects and allows us to add an object to the end, remove an object from the beginning and ﬁnd out how many objects are queued (getCount). Someone has already implemented a List class that allows us to add an object to the end, remove the ﬁrst object (using removeFirst) and ask how many objects there are (using getSize) – exactly the kind of messages needed for Queue, but with different names. We don’t want to force our clients to use the List class because it has the wrong name and messages and because the client would be able use it for non queue-like behavior (such as removing an object from the middle of the queue with removeElementAt). One solution, which avoids writing lots of code but still gives clients what they need, is to implement the queue object as an adapter (see Figure 11.21). When ﬁnished, our Queue translates its three messages into the corresponding List messages but omits all of the unwanted List behavior, as shown in the following Java implementation:

Common Design Patterns

Client

? Queue

List

+add(:Object) +remove():Object +getCount():int

+put(:int,:Object) +get(int):Object +remove(:int) +getSize():int ...

Figure 11.20: Queues and Lists

List

Queue -queue Client

+add(:Object) +remove():Object 1 +getCount():int

-list +put(:int,:Object) +get(int):Object 1 +remove(:int) +getSize():int ...

Figure 11.21: Queue class diagram public void add(Object o) { list.add(o); } public Object remove() { return list.removeFirst(); } public int getCount() { return list.getSize(); }

355

356

Chapter 11

11.3.9 Strategy and Template Method Deﬁne a family of algorithms, encapsulate each one, and make them interchangeable. Strategy lets the algorithm vary independently from clients that use it. [Gamma et al. 95] Sometimes, a family of tasks differ in some minor detail. For example, driving a car is similar to driving a van or a lorry, but each has its own detailed characteristics such as stopping distance, all-round visibility and driver height. To take a smaller example, how about the problem of keeping a list of objects in sorted order? We can make sure that the elements of a list remain sorted by inserting each new one into its correct location, just as we would do with physical index cards. To be more speciﬁc, when a new element comes along, we scan down the existing list to ﬁnd the ﬁrst element that is ‘bigger’ than the new one, then we insert the new one in front of the one that we just found. (If we get to the end of the list without ﬁnding a bigger element, we know that we have to add the new element at the end.) Although the general task is the same each time, the detail is different: if the list contains numbers, ‘bigger’ means ‘further from zero’; if the list contains names, ‘bigger’ means ‘further from A’. In software terms, we now have a problem: because the detail is different every time, we can’t easily implement a SortedList class that works for every kind of element. Here are a couple of things that we could try: • Make the add(:Object) method use an if statement to work out what to do (if o is a number, do this; otherwise, do that). The trouble with this is that the varying logic is buried inside a method, where it’s difﬁcult to ﬁnd. Worse still, clients can’t use our class to deal with lists of customers sorted by credit rating, or products sorted by price, or anything else they might think of. • Provide an abstract method called biggerThan(:Object,:Object) that returns true if the ﬁrst parameter is bigger than the second (this method would be invoked by a generic algorithm inside add). Then, we would need to provide a SortedNumberList subclass that implements biggerThan in one way and a SortedNameList subclass that implements it another way. (This is actually another pattern called Template Method.) Although the kernel of the sorting algorithm is now easy to get at, we’ve disturbed the collection hierarchy: what if SortedList needed subclasses called SortedLinkedList and SortedArrayList? We could end up with SortedNumberLinkedList, SortedNameLinkedList, SortedNumberArrayList, SortedNameArrayList – a complete mess. (Any alternative solution using multiple inheritance would probably not be any easier to use.) Strategy is designed to solve this problem. Essentially, we take the detail (the strategy) and separate it completely from the main task (the context). In this case, we separate the biggerThan method from the SortedList class, as shown in Figure 11.22. Here, SortedList

Common Design Patterns

SortedList +SortedList(:Comparator) <> +add(:Object)

* 1

-comparator

Comparator +biggerThan(:Object,:Object):boolean

NumberComparator

NameComparator

+biggerThan(:Object,:Object):boolean

+biggerThan(:Object,:Object):boolean

Figure 11.22: Class diagram for a SortedList using the Strategy pattern

has an add method as before, but now biggerThan is inside an interface called Comparator. Comparator represents an object that knows how to compare two values of a similar type: it has subclasses for each type of object that we’re interested in (NumberComparator and NameComparator). Each SortedList must have a Comparator available when objects are added. Therefore, we give SortedList a comparator attribute, set by the constructor. Now, if we want to build a sorted list of numbers we do the following (see the sequence of messages in Figure 11.23): Comparator c = new NumberComparator(); SortedList l = new SortedList(c); l.add(aNumber); l.add(anotherNumber);

If we want sorted names instead, all we have to do is: Comparator c = new NameComparator(); SortedList l = new SortedList(c); l.add(aName); l.add(anotherName);

The nice thing about Strategy is that it allows client programmers to implement their own strategies without touching the SortedList class at all. For example, if we want a sorted list of WeddingGift objects, we would implement our own WeddingGiftComparator with a biggerThan method that compares prices. It’s as if the context is an incomplete jigsaw to which anyone can add the last piece.

357

358

Chapter 11

aClient

aSortedList

aNumberComparator

add(num1)

add(num2)

biggerThan(num2, num1)

true

Figure 11.23: SortedList sequence diagram

Of course, the insertion sort described in our SortedList example is horribly slow. It would be much faster to jump to the middle of the list to begin with: if we hadn’t gone far enough, we would then jump half way towards the end; if we had gone too far, we would jump half way towards the beginning; we would repeat this process until we found the correct location. This algorithm is called a binary chop search. It was omitted from our earlier discussion because it would just have got in the way.

11.3.10 Flyweight Use sharing to support large numbers of ﬁne-grained objects efﬁciently. [Gamma et al. 95] Flyweight is a simple pattern that allows client programmers to think that they’re using a factory method to create their own object, when ‘their’ object is actually being shared by multiple clients. Normally, this is done to save memory and improve performance, by avoiding the creation of many equivalent objects. For example, if we wrote an application that dealt with people’s ﬁrst names, we would discover that many people had the same ﬁrst name. Why should we create hundreds of instances of the name ‘Sam’ when we could create just one object, with multiple references to it? The two alternatives are shown in Figure 11.24. What we need is a cache, or pool, of names that have already been created. Thus, rather than creating a name every time we needed one, we could look in the pool to see if one already existed; if it did, we could use it; if it didn’t, we could create a new one, add it to the pool and then use it (see Figure 11.25).

Common Design Patterns

:Name myName firstName = "Sam" myName :Name yourName

OR

:Name yourName firstName = "Sam"

firstName = "Sam" herName :Name herName firstName = "Sam"

Figure 11.24: To share or not to share

Ideally, we would like the complexities of searching the pool and creating a new object to be hidden from client programmers: a Factory Method is ideal for this. Figure 11.26 shows a generic Flyweight alongside one that is speciﬁc to the name example. Flyweights don’t have to be 100% shared: if we make the clients store some of the state externally and then pass it back in as a parameter to every ﬂyweight message, we have a hybrid. To take a trivial example, let’s assume that we want to add a person’s second initial to their ﬁrst name. If we added the second initial to the ﬂyweight itself, we would lose most of the sharing (‘Sam J.’ is far less common than ‘Sam’). However, if we arrange for the clients themselves to manage the second initials, as long as the second initial is passed to the ﬂyweight as a parameter to every message, the ﬂyweight will still have all the information it needs. For example, in the following Java code fragment, we’re performing some action if the ﬁrst name or the second initial contains the letter ‘j’: public class Name { ... public boolean containsLetter(char letter, char secondInitial) { String namePlusSecondInitial = givenName + secondInitial; return namePlusSecondInitial.containsIgnoringCase(letter); }... } Name n = aNameFactory.createName("Sam"); char secondInitial = 'J'; if (n.containsLetter('j', secondInitial)) { ...

State inside a Flyweight is referred to as intrinsic, while state managed externally, by the client, is extrinsic. The Flyweight pattern is particularly appropriate, and easiest to implement, when the ﬂyweight’s state is immutable (read-only), otherwise we tend to get

359

360

Chapter 11

myName herName

:Name

POOL

givenName = "Sam"

:Name

:Name

givenName = "Mike"

givenName = "Fred"

n

name2

:Name manager

givenName = "Sue"

yourName

Figure 11.25: Sharing objects in a pool

FlyweightFactory available to client (e.g. as Singleton)

NameFactory

+createName():Name 1 FlyweightFactory

*

+createFlyweight():Flyweight

-pool: Collection

Name

1

* Client

-pool: Collection

Client

-firstName: String ~Name(:String) <> +getFirstName():String

Flyweight

Generic

Specific

Figure 11.26: Flyweight class diagrams

Common Design Patterns

more extrinsic state and less sharing. For example, consider changing the second character of the name ‘Sam’ to ‘i’; would we really want to affect all other uses of the name?

11.3.11 Composite Compose objects into tree structures to represent part–whole hierarchies. Composite lets clients treat individual objects and compositions of objects uniformly. [Gamma et al. 95] A part–whole hierarchy is another name for a composition or an aggregation. Despite its name, Composite doesn’t describe a composition in the strict UML sense; instead, it describes a design that we can use to implement an aggregation or a composition. Thus, Composite allows us to build a hierarchy of objects with an arbitrary number of levels. As an added bonus, Composite allows us to treat all levels of the hierarchy in the same way – this would allow us to, say, remove an entire section just as easily as we could remove a single piece. Figure 11.27 shows one popular form of Composite. Here, we have an abstract class called Component that represents part of the hierarchical structure. Underneath that, we have the class Basic that represents the smallest possible component (one that can’t contain any other components). Alongside Basic is the Composite class that represents a component that can contain other Components (hence the link back to Component). Because Composite can contain any type of Component – Composite or Basic – we can build any number of levels, rather than just two. The use of aggregation with a * multiplicity may look a little odd at ﬁrst, but there are three good reasons: ﬁrstly, a Basic component doesn’t have to be inside a Composite at all (although it often is); secondly, although it would be a rare situation, the pattern itself doesn’t prevent us from putting a Component inside more than one Composite at a time (if it did, the relationship might have been a UML composition); thirdly, the pattern doesn’t specify that a Component should die when its Composite dies (although it often will) – shared death is another requirement for UML composition. Because Composite has a link back up the hierarchy, it is often referred to as recursive (as in recursive composition). Recursive, in computing terms, means going around again and again. The end result is a bit like a Russian doll: the biggest doll has a smaller doll inside it, which has a smaller doll inside it, which has a smaller doll inside it, until, eventually, we get to the smallest, basic, doll. Unlike the Russian doll, however, each Composite can contain any number of components – this allows us to build broad hierarchies, with any number of children below each parent (three ﬂoors in a house, ﬁve rooms on each ﬂoor, four walls in each room).

361

362

Chapter 11

Component

-children:List

Client * +message()

* Composite

Basic

+message()

+message() +add(:Component) +remove(:Component) +createlterator():lterator +createTwoWaylterator():twoWaylterator

Send message() to every child

Figure 11.27: Composite class diagram

Figure 11.28 shows two hierarchies described by the generic Composite model. Here’s how we might use Java to build hierarchy A: Component small = new Component(), medium = new Component(), large = new Component(); Basic atom1 = new Atom(), atom2 = new Atom(), atom3 = new Atom(), atom4 = new Atom(); small.add(atom1); small.add(atom2); medium.add(small); medium.add(atom3); large.add(atom4); large.add(medium);

The children in the Composite examples are ordered, left to right: generally, it’s a good idea to preserve the order in which components are added, in case it’s important to the client programmer. A createIterator method is provided which returns an instance of the Iterator pattern. For completeness, a createTwoWayIterator method is provided, to allow clients to iterate backwards.

Common Design Patterns

ourHouse large

... upstairs medium

atom4

... suesRoom atom3

small

wall3

wall1 atom1

atom2 wall4

wall2

A

B Figure 11.28: Composite object diagrams

Apart from dealing with the hierarchical structure, Composite also ‘lets clients treat individual objects and compositions of objects uniformly’. In design terms, this means that any message at the Component level will also appear at the Basic and Composite levels, as is the case with message in Figure 11.27. A surprising (and pleasing) effect of the recursive nature of Composite, is that we can simply get message on Composite to pass the message on to its children. To see how this works, consider the house example in Figure 11.28B. If pink were our favorite color, we could paint every wall in the house pink. To do this, we would send the message paint("Pink") to ourHouse; ourHouse would send paint("Pink") to each ﬂoor in turn, up to and including upstairs; the ﬂoors would send paint("Pink") to the rooms; ﬁnally, paint("Pink") would be sent by the rooms to the walls; since the walls are basic objects, they would do the actual work. Ultimately, a single message sent by the client programmer results in every wall in the house turning pink. (To convince yourself of this, try drawing the messages on a sequence diagram.) You may be wondering at this point what the difference is between collections and composites, especially since iterators have cropped up in our discussion. While it’s true that collections and composites both contain objects, it turns out that most

363

364

Chapter 11

clients of collection classes are happy with four basic classes (and maybe sorted versions too): • • • •

Bag:

Unordered objects with duplicates (not available in Java). Set: Unordered objects without duplicates. List: Objects in a speciﬁed order. Map: A lookup table that needs a key to add or remove an object.

The common factor between these classes is that client programmers don’t have to build the structure; they simply add and remove objects. (Okay, in the case of List, they do have to specify the position of each object but they don’t have to create the object holders or arrange them side by side.) The Composite class, in contrast, forces the client programmer to build the structure piece by piece, explicitly combining small pieces into larger ones to produce some tree-like structure. In practice, we tend to use the Composite pattern if the client wants to or must handle the structure explicitly (because it could not have been determined beforehand) or if the client wants the convenience of a top-level message that gets sent to all children. (Can you think of a way of using Strategy to allow collection classes to send a top-level message to all children?) Composite is sometimes used as part of the internal implementation of collection classes, to build an index, for example. Another area where you may encounter the Composite pattern is in user interfaces, since GUIs are usually treated as hierarchies of components.

11.3.12 Proxy Provide a surrogate or placeholder for another object to control access to it. [Gamma et al. 95] Proxy involves one object (the proxy) interposing itself between a client and another object (the real subject), as illustrated in the lower diagram in Figure 11.29. We might do this for any number of reasons, such as security control, lazy initialization or remote access. In the example illustrated, clients send message to the proxy which passes the message on to the real subject (which might involve checking security permissions, creating the real subject the ﬁrst time around, or performing the network communication). Although not strictly necessary, it’s a good idea to introduce an interface – Subject, for example – that lists the messages appearing on RealSubject and Proxy. For one thing, this makes sure that the interfaces of the proxy and the real subject are kept in step; secondly, it’s reasonable from a modeling point of view, because the proxy and the real subject do the same thing.

Common Design Patterns

Subject Client

-subject +message()

Proxy

RealSubject -realSubject

+message()

*

+message()

message()

message() aClient

1

aProxy

aRealSubject

Figure 11.29: Proxy class and object diagrams

Probably the most common use of Proxy is in the area of network programming, which involves one piece of code invoking another piece of code that resides on a separate machine, with some kind of network between them. In the ﬁeld of object technology, this mechanism has become a well-known part of frameworks such as CORBA, J2EE and .Net. In object-oriented programs, the only way we can run a piece of code is to send a message to an object. But how do we send a message to an object that lives on a different machine? The answer is that we provide a local proxy that has the same interface as the remote object but whose methods perform the network communication for us. This way, the client is shielded from the complexities of the network communication. Using a proxy to send a remote message is illustrated in Figure 11.30 (which combines elements of a deployment diagram and a communication diagram). Here, we have a remote system for looking up an employee’s e-mail address by payroll number. The server machine has a process containing a ContactsImpl, called s, which waits for incoming client requests. The client machine has a process containing a ContactsClient, called c, and a ContactsProxy, called p, with the same interface as s. When c wants an e-mail address, it sends lookup to p, which does whatever it takes to pass the message on to s (create a communication channel, identify the type of message, pass on the parameters, and wait for the response). On the

365

366

Chapter 11

:ClientProcess

:ServerProcess

c:Contacts Client p(...)

looku

lookup(...)

s:Contacts Impl

mmun

rk Co

Netwo

Link ication

p:ContactsProxy

Figure 11.30: Remote object implemented using Proxy

server side, some complementary code (not shown) receives the incoming request and passes it on to s as a normal message. When s receives the message, it executes its lookup method, in exactly the same way as if the message had originated locally. When s has ﬁnished, the reply is returned to p. On receiving the reply, p completes its lookup method and returns the result to c. The classes and interfaces involved in this proxy example are shown in Figure 11.31. In this ﬁgure, unlike most design-level class diagrams, the association between the ContactsClient and the ContactsImpl is not navigable. This is because, in the case of a remote object, the proxy doesn’t have a physical pointer to the real subject: it has a reference to the communication channel, or to some kind of address for the subject (if it doesn’t want to keep the communication channel open all the time).

Contacts Contacts Client

+lookup(payroll:String):String

ContactsProxy

ContactsImpl realSubject

+lookup(payroll:String):String

*

1

+lookup(payroll:String):String

Figure 11.31: Remote object class diagram

Discovering, Combining and Adapting Patterns

11.4 USING PATTERNS As you use patterns more and more, you will ﬁnd that they become second nature for particular tasks, especially well-understood tasks such as creating objects (factories), remote objects (proxies) and complex life cycles (states). For other tasks, you have to watch out for areas where they might be useful. You will also ﬁnd yourself using several patterns together to accomplish a single task (we saw an example of this when Factory Method was used as part of Iterator).

11.5 DISCOVERING, COMBINING AND ADAPTING PATTERNS Although there is no deﬁnitive advice on how to do discover and combine patterns, it’s a good idea to start by practicing with the patterns outlined in this chapter. Then, look for how patterns have been used in the code around you, in libraries and frameworks and in code written by your colleagues. Be sure to keep returning to [Gamma et al. 95] and the other resources mentioned in this chapter in order to expand your knowledge. There’s a lot to absorb but the end result is better productivity, higher quality code and more peace of mind. One thing that you will discover quickly is that patterns need to be adapted to each particular use. Here are some of the reasons for this: • Language translation: The pattern you started with may not be implemented in your chosen programming language. All programming languages, even object-oriented ones, have differences. Some differences are large (such as how pure or hybrid the language is) and some are small (for example, the access protections available for messages), but the standard patterns will have to be modiﬁed to suit each language. • Difference of opinion: For each design problem, there is more than one solution. Just because a pattern has been produced by an expert, or a group of experts, doesn’t mean that it’s perfect. Once you fully understand a pattern, feel free to change it, especially as you become more expert. • Difference in style: Design and coding styles can come from many sources: experience, company coding guidelines, requirements of a library or framework, and so on. If a pattern doesn’t ﬁt the style that your project uses, adapt it. (Often, doing things the same way as those around you is more important than doing things in a better way – consider the beneﬁts of reuse and easier maintenance.) • Different problem: Each pattern is as general as possible – some of the class models presented here would never be used as they are. For example, in the Observer pattern, we had the Observer, Subject, ConcreteObserver and ConcreteSubject classes; we would not

367

368

Chapter 11

expect to see any real application with classes called ConcreteObserver and ConcreteSubject. No pattern will ﬁt your problem exactly. • Composition, inheritance and multiple inheritance: Languages, and developers, differ in the way they use inheritance and composition. You may ﬁnd yourself performing transformations from one representation to another, while trying to keep the overall effect of the pattern the same. • Consequences: Every good pattern description includes consequences, areas where the author has identiﬁed particular beneﬁts, shortcomings and trade-offs of using the pattern. (No pattern is perfect, after all.) As a result of this advice, you may decide that you prefer different trade-offs. • Clarity: Sometimes, a full-blown pattern with all the extras can be more complex than you require. A little simpliﬁcation can be beneﬁcial (especially in a learning context). All of these factors can be observed to some extent in the pattern descriptions given in this chapter. Language translation, at least, was necessary because the notation used here is UML and the programming language is Java, rather than the combination of OMT and C++/Smalltalk used in [Gamma et al. 95]. Although the Observer pattern is undoubtedly a good idea, it has certain drawbacks in its simplest form – Figure 11.32 shows an adaption of Observer, reworked to deal with these shortcomings. • Whenever there is a change in the state of the subject, update is called. An observer may read attributes that haven’t changed and refresh itself unnecessarily. In the adapted pattern, the client speciﬁes during registration which attribute the Observer is interested in, using the new parameter a (of course, the Observer can be registered for more than one attribute). The notify method also takes an a parameter: it only notiﬁes Observer objects that have been registered for changes to a. Observers can observe some attributes and ignore others. • The observer has to return to the subject to ﬁnd out what the changes were, which involves more communication. In the adapted pattern, the update method takes the new value of the attribute as a parameter. Thus, the Observer doesn’t have to return to the Subject to use a getter. • The observer can observe only one subject. In the adapted pattern, the Subject is passed as a parameter to update. This allows the Observer to be registered with any number of subjects while still being able to identify them. When using this modiﬁed form of Observer, the subject must notify observers when derived attributes change, not just stored attributes.

Discovering, Combining and Adapting Patterns

Subject +addObserver(a:String,o:Observer) +removeObserver(a:String,o:Observer) #notify(a:String,v:Object)

<> Observer

-observers

*

*

+update(s:Object,a:String,v:Object)

ConcreteSubject ConcreteObserver

-attribute:X +getAttribute():X +setAttribute(v:X)

Set attribute and call notify("x",v)

+update(s:Object,a:String,v:Object)

Parameter names: v is value a is attribute name o is observer s is subject

Figure 11.32: Class diagram for an improved Observer pattern

Case Study Patterns in iCoot Let’s see what impact the patterns described have on the iCoot system: • Observer is used on the client side by the GUI interfaces (both desktop and mobiledevice versions). These interfaces use a two-layer model consisting of GUILayer and ControlLayer – all upward communication uses Observer in its Java form, event delegation. In event delegation, all notiﬁcations take the form of events broadcast to listeners. Some events signal a change in attribute values and some, pure events such as ‘logon succeeded’, do not. • Singleton is used on the server side where one home per entity manages the creation and ﬁnding of entities. Some homes also have general utility methods such as carModelHome.ﬁndEngineSizes(). • Multiton is not used. Despite the recent addition of multitons to the syntax of Java, it is still common to use public class constants of type int as an alternative, as in CarColor.MIDNIGHT BLUE. (A class constant is a read-only (i.e. ﬁnal) class ﬁeld.) • Iterator is used on the client and the server for all access to List objects.

369

370

Chapter 11

Case Study (cont’d) • Factory Method is used widely on the client and the server. For example, when the business layer needs to create an entity that contains data from the database, a factory method on the home is used: CustomerHome.getInstance().create(”Fred Bloggs”, ”02723359853”, 0). • Abstract Factory is not used, although, in principle the homes might inherit from a common superclass. • State is used to represent the state machine for a Reservation. • Facade is used to encapsulate layers. The cleanest example is the ServerLayer objects which hide the complexity of the BusinessLayer from clients. • Adapter is used in the Java-based clients to connect GUI components, which expect to send messages to event listeners, with controller objects, which have a generic interface. • In the next increment, Strategy will be used in user interfaces to allow retrieved cars to be sorted according to attributes selected by the user. • Template Method is used wherever possible in abstract classes to share logic between subclasses (to improve the quality of code, make maintenance easier and reduce the effort required). • Flyweight is used on the server, where the business layer maintains a cache of entities populated with data from the database. Flyweight is used to avoid duplication of the data within the server process and to prevent different parts of the application working with different versions of the same data. • Proxy is used in the Java-based clients. The ControlLayer uses proxies to access the facades in the ServerLayer (passing lightweight copies backwards and forwards). • Composite is used in the GUI clients. It is part of the standard Swing facilities for constructing component trees.

11.6 SUMMARY In this chapter, we looked at: • How software developers use design patterns to avoid duplicating effort by applying the knowledge and experience of other developers to a particular problem. A pattern describes a particular way of doing something that has proved effective in real world projects. • How patterns are used to record knowledge, to document solutions and to aid reuse.

Further Reading

• The pattern template proposed in [Gamma et al. 95]. • 14 of the most important patterns, to provide you with a good grounding and to whet your appetite. • The practical implications of using patterns: how to discover and combine patterns and how to adapt them to your particular problem.

FURTHER READING The original Gang of Four book [Gamma et al. 95] is invaluable reference material despite the fact that the book is illustrated using OMT, C++ and Smalltalk. It contains a lot of extra discussion and advice. Since the publication of the original book, many more books and research papers have been produced, and many more areas of interest have been explored. For further reading on pattern history and the state of the art, try the following Web sites: www.hillside.net and www.patterndigest.com. For a comprehensive description of how to implement patterns in Java, see [Grand 02] and [Grand 99].

371

12

Specifying the Interfaces of Classes

I

n this chapter, we consider the code speciﬁcation process, traditionally the last development phase before implementation. Speciﬁcation allows us to write down how our software should behave in a way that is less ambiguous and more precise than anything we’ve used so far.

Learning Objectives Understand what a speciﬁcation is and why we might want to produce one.

•

Understand the distinction between formal and informal speciﬁcations.

Write speciﬁcations in an object-oriented manner.

•

Understand the Design by Contract philosophy.

•

Use speciﬁcations and Design by Contract in Java.

Chapter Outline 12.1 Introduction 12.2 What Is a Speciﬁcation? 12.3 Formal Speciﬁcation 12.4 Informal Speciﬁcation 12.5 Dynamic Checking 12.6 Object-Oriented Speciﬁcation 12.7 Design by Contract 12.8 Informal Speciﬁcation in Java 12.9 Summary Further Reading Review Questions Answers to Review Questions

12

374

Chapter 12

12.1 INTRODUCTION Although, in one sense, we’ve been specifying the behavior of our software at all the stages (from system use cases all the way through to the database schema), the discussion here relates to the interfaces of the system components, and classes in particular. So why do we write a speciﬁcation? Here are some good reasons: • To remove the ambiguity remaining after earlier phases: Ambiguity means that artifacts can be read more than one way. Use cases are typically written in natural language, which is notoriously vague; analysis is not meant to dictate to the designer how the code should actually be implemented, so the analysis artifacts don’t match the ﬁnal code exactly; as for design, we may have speciﬁed exactly which classes are to be written and what their messages and ﬁelds should be but there may be no documentation of how the messages should work together or what constitutes a legal message parameter, a legal ﬁeld value or other such detail. • To improve our understanding of the implementation: The more we understand our design, the easier it will be to implement. Writing a speciﬁcation requires us to think hard about our design, so we end up with a better understanding. • To increase our conﬁdence that the system will work: Often, in attempting to write down exactly how our software should behave, we discover that we’ve missed something out or assumed something that’s impossible, or made some other human error. Thus, speciﬁcation allows us to spot mistakes sooner than if we plough straight on into implementation. • To help us debug our software: When our software misbehaves, we can check whether the implementation matches the speciﬁcation. Thus, we have a starting point for identifying and removing faults. • To help us test our software: A speciﬁcation describes how our software must behave and a test veriﬁes that our software behaves as described. Thus, tests can be used to check that our software ﬁts the speciﬁcation. Tests based on system use cases verify that we have produced a system that meets the system requirements. Speciﬁcation-based testing is essential for libraries, patterns, frameworks and reusable objects, because none of these involve an actual system. • To help us modify our software: If we decide to modify our system in some way – by adding functionality, for example – we can check that the modiﬁcation does not break the speciﬁcation. If the speciﬁcation would be broken, the proposed modiﬁcation is incompatible with the current system (or we have an error in the speciﬁcation). Rather than just go ahead and break the speciﬁcation, we should look at the proposed change more carefully.

What Is a Speciﬁcation?

• To allow us to pass the implementation task on to other developers: Although the other developers are still going to have to contact us to clarify detail, they should need to do so less often. • To provide better documentation: Better documentation leads to easier maintenance, more reuse and less misuse.

12.2 WHAT IS A SPECIFICATION? In general terms, a speciﬁcation is a complete, unambiguous description of the required behavior of a piece of software. The piece of software in question could be an entire system, a subsystem, a layer, a class or a function. Although we would like our speciﬁcation to be complete and unambiguous, that’s not always practical or possible. However, even imperfect speciﬁcations have their uses (and they’re better than nothing). A speciﬁcation describes one or more boundaries. For example, the boundary of a function library consists of the signatures (names, return types and parameter types) of its functions and of its global data values; the boundary of an object consists of its messages and its attributes; the boundary of a class (subtly different to the boundary of an object) consists of its class messages and class attributes. Although attributes logically sit on the boundary, we expect them to be accessed by object and class messages. If we consider that each boundary has a client on one side and a supplier on the other, a speciﬁcation describes: • What information the client can pass over the boundary and when they’re allowed to pass it; for example, the client may only be able to pass new address information to a customer object once a transaction has been started. • What information the client can retrieve from the boundary and when they’re allowed to retrieve it; for example, the client may only be able to retrieve a list of books for sale from a book server once it has logged on. • When an event will be broadcast to registered clients and what information the clients will receive; for example, in a distributed chat system, every client will be notiﬁed of the name of each new client that joins the chat room. • The legal states of the information managed by the boundary; for example, a collection object may state that it will never contain a negative number of objects. Most of the time, a speciﬁcation is concerned with the public boundary of a piece of software. However, we can equally well specify the nature of internal boundaries. As an

375

376

Chapter 12

example, consider a function library that provides sorting facilities for collection structures. Such a library might have a reusable swap function for swapping two values in a collection. It’s unlikely that the swap function would be part of the public interface of the sorting library because it’s an implementation detail; however, we could still specify its behavior – we could state, for example, that after the function has been called, the left-hand value will occupy the location previously occupied by the right-hand value, and vice versa. There are two varieties of speciﬁcation: a formal speciﬁcation is scientiﬁc and rigorous; an informal speciﬁcation is pragmatic and partial, but no less useful.

12.3 FORMAL SPECIFICATION For the most part, software development is an imprecise activity. Historically, software has been developed more in the hope than in the certainty that it will do what it’s supposed to do. This is because there is no science to back it up, unlike disciplines such as civil engineering or hardware engineering, which are backed by mechanics and Quantum theory. Most good software is a result of human ingenuity, experience, guess work, reuse and testing. This is ﬁne for noncritical applications such as word processors and consumer operating systems. We tolerate a word processor occasionally crashing in the middle of an edit (this is why we have features such as auto-save). We expect a consumer operating system to crash sometimes, leaving us with no alternative but to reboot (which is why system administrators run automated backups). We may lose time, we may get frustrated and angry, but no-one’s life is endangered. The advantages of being able to use computers at all generally outweigh the disadvantages of those computers being imperfect: on the whole, we’re more productive than if we had no computers at all. However, some situations call for perfect, or near-perfect, reliability. Safety-critical systems are a good example – these include the software at the heart of nuclear power stations, air trafﬁc control systems, ﬂy-by-wire aircraft and hospital equipment. We can’t afford to let such systems (or their operating systems) fail. In reality, we can’t guarantee 100% reliability, because that would be too difﬁcult or too expensive or would require knowledge of every possible abnormal condition. Instead, we settle for either a very long time between failures (for example, a control system in a nuclear power station may be acceptable if it makes a mistake, on average, less than once in a trillion years) or better reliability than a human (for example, a ﬂy-by-wire plane may be acceptable if it makes one-tenth of the mistakes of a human pilot). One way to improve reliability is to replicate hardware or software. For example, we could install three computers to control a ﬂy-by-wire plane. Each computer would use different hardware, a different operating system and software developed by a different team. We could arrange these three computers in a voting system, where any command would be ignored by

Formal Speciﬁcation

the plane unless it came simultaneously from at least two of the computers. Theory suggests that, using such techniques, we can make failure very unlikely. (We still can’t be perfect, because two of the computers could fail in the same way at the same time, causing the plane to follow an incorrect command.) Further discussion of this kind of reliability assurance is beyond the scope of this book. Another way to improve reliability involves trying to prove, in the mathematical sense, that the software is correct (that it does what it’s supposed to do). This involves three steps: 1. Produce a formal speciﬁcation, in a mathematical language, that describes how the software should behave. 2. Prove that the speciﬁcation is feasible – i.e. that it has no logical contradictions or impossibilities. 3. Prove that the software conforms to the speciﬁcation. Examples of formal speciﬁcation languages are Vienna Development Method (VDM) (originally from IBM Vienna), Z (originally from Oxford University) and Object Constraint Language (OCL) [OMG 03b], which is part of UML. For example, if we want to write a formal speciﬁcation for the behavior of a squareRoot() function, here are some things that we might want to say about it: • The input value must be positive. (Assuming noncomplex numbers.) • The square of the result will be equal to the input value. (This is a common trick when specifying a reversible function: applying the inverse function to the result must yield the input value.) • The result will be positive. (Remember that (2 × 2 = 4) and also (−2 × −2 = 4), so 4 has two square roots; we choose to return the positive square root.) Here’s how we can specify the boundary conditions of squareRoot() in VDM: squareRoot(x : R) y : R pre x ≥ 0 post (y2 = x) ∧ (y ≥ 0) where R stands for real numbers, pre stands for precondition (something that must be true before the function can be invoked) and post stands for postcondition (something that will be true after the function has been invoked). Every logical expression in a speciﬁcation is referred to as an assertion (so y ≥ 0 asserts that y will be greater than or equal to 0). Even if you do not know VDM, it’s easy to see that the speciﬁcation above reads as: The squareRoot() function takes a real number x as parameter and returns a real number y. For correct operation, the function requires that x is greater than or equal to 0. If

377

378

Chapter 12

the precondition is met, the function guarantees that y squared will be equal to x and that y will be greater than or equal to 0. Thus, we can specify the exact behavior of squareRoot() without saying anything about its implementation. Formal speciﬁcation is a discipline that requires a great deal of training and skill. It is also a lengthy process. Sometimes, it’s possible to use a computer to validate a speciﬁcation (automated theorem proving). Under very limited conditions, we may even be able to verify that an implementation matches the speciﬁcation. However, no computer could write the speciﬁcation itself. Even if you can rely on the mathematics of the formal speciﬁcation and on your software implementation, you can’t rely on the operating system, the hardware or the compiler. For this reason, safety-critical software is sometimes implemented in assembly language and then deployed on a cut-down operating system, so that we have a chance of proving that the end result is correct. (Even this doesn’t stop cosmic rays ﬁring through a chip and ﬂipping a bit from 1 to 0, so we need external fail-safes too.) We also can’t prove that the formal speciﬁcation meets the system requirements, especially since the latter usually start out as natural language and change subtly over time. Because formal speciﬁcations are so difﬁcult and time-consuming to produce and to use, they’re normally only employed for safety-critical systems and those that require advanced reliability techniques, such as voting systems. Although formal speciﬁcations are impractical for the software industry as a whole, we can still apply the underlying principles and get many of the beneﬁts. Colloquially, the techniques described in the rest of this chapter are referred to as informal speciﬁcation. With the help of formal speciﬁcation theory, we can at least be clear about what an informal speciﬁcation should contain.

12.4 INFORMAL SPECIFICATION All programmers use informal speciﬁcations to a certain extent – adding a comment to a function for the beneﬁt of other programmers is a simple example. A function comment will describe some or all of the information in the following list: • • • • • •

when the client can call the function what parameters should be passed what the function does what kind of result is returned (type and value) what effect the function has on global data what action the function takes if there is a problem

Informal Speciﬁcation

Such information applies equally to subroutines, procedures and methods, since these are just variations on the basic idea of a function. Usually, a function comment won’t say anything about the internal implementation – just like any other form of speciﬁcation, the comment only describes the conditions at the boundary. Occasionally a comment will say something about the implementation, but only when necessary and only in abstract terms. For example, it would be useful for a client programmer to be told the time/space trade-off of a particular sort function, such as This function executes at n-log-n speed and requires 2n internal locations for processing. The following piece of C code (C looks a lot like Java) contains a comment that is used as an informal speciﬁcation of the squareRoot function: /* Returns the positive square root of x. Preconditions: x >= 0. */ float squareRoot(float x);

Anyone reading this comment knows that they will get back the positive square root of x as long as they provide a positive x. The term precondition should be familiar to most programmers, or easy for them to work out. If you’re concerned about maximum readability, you could use Requires that: instead of Preconditions: and Ensures that: instead of Postconditions:; or you could write the preconditions and postconditions in natural language. The informal speciﬁcation above is incomplete, unlike the formal VDM version that we saw earlier: for example, we have said nothing about the square of the result being equal to the parameter. In the next section, we will see why this particular postcondition is not as straightforward in programming as it may seem. A function signature (name, return type, parameter types) contains a lot of information in its own right so it could form part of the informal speciﬁcation. For example, ﬂoat squareRoot(ﬂoat x) implies that the function takes a ﬂoat value and returns a ﬂoat value; the result will be the square root of the parameter. However, even an informal speciﬁcation should not rely on implied information, hence the explicit comment. Since comments and signatures can be included in a number of UML artifacts, source code is not the only place where informal speciﬁcations can appear. Another kind of informal speciﬁcation that we’ve encountered in this book is the use case. Each use case describes some or all of the following information: • • • •

when the use case can be used (preconditions) what the use case does (steps and postconditions) what effect the use case has on the system (steps and postconditions) what happens in abnormal cases

379

380

Chapter 12

Thus, since use cases describe everything an external actor needs to know in order to use the system properly, they’re similar to function comments. This should be no surprise since, by deﬁnition, a use case is supposed to describe a boundary – the system boundary in this case. For the purposes of this chapter, we’re interested in the speciﬁcation that takes place between subsystem design and implementation, so let’s leave use cases for now. Some programming languages, notably Eiffel, have a special syntax for recording informal speciﬁcations in source code, separate from comments.

12.5 DYNAMIC CHECKING Authors of formal speciﬁcations use a mathematical notation that is separate from the software implementation. A formal speciﬁer may be able to prove that the speciﬁcation they produce has no logical errors. In limited circumstances, they may even be able to prove that a particular software implementation conforms to the speciﬁcation. But, generally, formal speciﬁers rely on programmers using their knowledge and intuition to produce code that conforms to the speciﬁcation. The informal approach, in contrast, includes a pragmatic technique that we can call dynamic checking. A dynamic check is code embedded in our implementation that veriﬁes that the software is behaving itself (i.e. that it’s not breaking the speciﬁcation). To see how a dynamic check might work, consider the square root function again. Below is a C implementation of this function (with the calculations omitted): /* Returns the positive square root of x. Precondition: x >= 0. */ float squareRoot(float x) { if (x < 0) { fail("squareRoot", "Parameter x can’t be negative"); } ... /* Code to calculate y */ return y; }

Inside this function, we can see a piece of code that checks that the precondition x >= 0 has been met – if it hasn’t, the program gives up. For convenience, a fail function has been provided that prints the function name and the error message and then stops the program (using the C library function exit). (Some languages provide a more elegant way of signaling failure using exception handling, but the facilities vary enormously.) The client programmer is now protected from themselves: most of the time, the client calls squareRoot with a positive number and the function quietly succeeds; however, on the

Dynamic Checking

rare occasion that the client calls squareRoot with a negative number, it is told that the program can’t continue until the problem is ﬁxed. You may be thinking that our squareRoot function is performing ordinary error-checking. It is certainly doing something that programmers frequently do, regardless of whether or not they understand the speciﬁcation process. The interesting point, as far as we’re concerned, is that this piece of error-checking code has only been put there in order to verify that the client programmer has not broken the speciﬁcation – if they do break the speciﬁcation, they have a fault in their code. Dynamic checking only makes sense for informal speciﬁcations, not formal ones. There are many reasons for this, some of which are listed below: • The aim of formal speciﬁcation is to guarantee that the implementation doesn’t violate the speciﬁcation. With dynamic checking, in contrast, we accept that implementations tend to contain faults, so we look for violations at run time. • Some formal requirements can’t feasibly be checked. The fact that we can’t check something is no reason to leave it out of a formal speciﬁcation. • Some formal requirements can’t be expressed using imperative code. (An imperative programming language, by far the most common type, is one where the programmer has to tell the computer exactly what to do; with a declarative programming language, the programmer simply states what the outcome should be.) Again, such requirements should still be included in the formal speciﬁcation. • Some formal requirements can’t be expressed accurately as imperative code. For example, in the squareRoot case, we would like to specify that the result returned by the function really is the square root. One way to do this is to specify that the square of the result is equal to the parameter, as in For result y, (y * y) == x. But imperative arithmetic is imprecise: squareRoot(4.0) probably returns 2.0, but squareRoot(3.79512) might return 1.94811, a number that is not exactly correct. To deal with inaccuracy, how could we say that a result will be close to some value? Well, there’s a whole branch of computing called numerical methods devoted to this kind of problem. To apply numerical methods, we would need to calculate a maximum possible error and use that in our speciﬁcation. In the squareRoot example, if we encapsulate the maximum possible error inside a function called squareRootError and have access to a pos function that returns the positive value of its parameter, we can specify For result y, pos((y * y) - x) <= squareRootError(). This speciﬁcation can be added to the function comment and encoded as a check at the end: /* Returns the positive square root of x. Preconditions: x >= 0. Postconditions: For result y,

381

382

Chapter 12

(y >= 0) and pos((y * y) - x) <= squareRootError()). */ float squareRoot(float x) { if (x < 0) { fail("squareRoot", "Parameter x can’t be negative"); } ... /* Code to calculate y */ if (y < 0) { fail("squareRoot", "Calculation gave negative result"); } if (pos((y * y) - x) > squareRootError()) { fail("squareRoot", "Calculation was inaccurate"); } return y; }

Now, rather than just protecting the client programmer from faults in their own code, we’re guarding against faults in our code too. If we aren’t able to calculate a value for the maximum possible error (or if we decide that it is too much trouble), we should still add the speciﬁcation to the comment as natural language: ... The square of the result is equal to x, subject to the accuracy of the current platform. */

12.6 OBJECT-ORIENTED SPECIFICATION Formal speciﬁcation languages such as VDM were originally designed to be used with structured software development. Some of these languages have since been extended to work with object-oriented languages. Others, such as OCL, have been designed speciﬁcally for object-oriented software development. The main difference between a structured notation and an object-oriented notation is that the latter uses any number of class scopes rather than a single global scope. (The term scope means name space, partition, subarea, or anything else that you can put a fence around.) An object-oriented speciﬁcation is simpler than a procedural speciﬁcation, because all rules appear alongside their related concept (the class or object in question). An object-oriented speciﬁcation language must allow us to assert: • when a message can legally be sent to an object (message preconditions, expressed in terms of public attributes)

Object-Oriented Speciﬁcation

• the valid parameters for each message (message preconditions) • the effect a message has on the receiving object (message postconditions, expressed in terms of public attributes) • the valid replies from each message (message postconditions) • the conditions that are always met by the object (class invariants, conditions that will always be true for instances of the class, expressed in terms of public attributes) As well as objects, object attributes and object messages, we also have classes, class attributes and class messages; thus, the ﬁve categories above can also be applied to classes, although the notation usually makes it difﬁcult to tell the difference. As an example, consider a Container class with an add message. Let’s assume that we want to assert the following: • Precondition: An object reference passed to the add message must not be null. • Precondition: An object passed to the add message must not already be in the Container. • Postcondition: After the add message, the Container will have one more object than it had before. • Postcondition: After the add message, the Container will contain the object that was passed in as a parameter. • Invariant: The Container always contains a positive number of objects. In the next two sections, we’ll see how to express the Container assertions in OCL, a formal speciﬁcation language, and in Eiffel, an imperative programming language that includes syntax for informal speciﬁcations.

12.6.1 Formal Speciﬁcation in OCL OCL is UML’s formal speciﬁcation language, so it ﬁts well with object-oriented methodologies that use UML as their notation. For the Container example, we could use the following fragment of OCL (in which the contains method returns true if and only if the receiver contains o): context Container:: add(o:Object) pre: (o <> null) and not contains(o) post: contains(o) and (size = size@pre + 1) context Container:: inv: size >= 0

OCL allows us to connect assertions to a class using the context keyword, which means that we can also refer to global functions and global data in our speciﬁcations, for the sake of

383

384

Chapter 12

hybrid languages such as C++. Postconditions often need to refer to the value of an attribute before the method is executed – this allows us to talk about the effect that a method has on the receiver’s attributes. In this case, size@pre refers to the value of size before add is executed. The OCL fragment above can be read as: When sending the add message to a Container, the parameter o must be non-null and o must not already be in the container. Once the method has completed, the container will contain o and the size of the container will be one more than it was before. The size of a Container is always greater than or equal to 0.

12.6.2 Informal Speciﬁcation in Eiffel Eiffel, a language developed by Bertrand Meyer in the mid-1980s [Meyer 90], was designed with informal, dynamically-checkable speciﬁcations in mind. Eiffel programmers can put assertions in the source code itself, but separate from the comments and the implementation code: the assertions are recognized by the Eiffel compiler as part of the syntax. Eiffel uses require instead of pre, ensure instead of post and old instead of @pre. For the Container example, we could use the following fragment of Eiffel: class CONTAINER feature {ANY} -- Public stuff follows size: INTEGER add(o:OBJECT) is require (o != void) and not contains(o) do ... ensure contains(o) and (size = old size + 1) end invariant size >= 0 end

Eiffel’s mixing of speciﬁcation and implementation is convenient for the programmer. In addition, since the compiler understands the assertion syntax, the assertions can be veriﬁed (to make sure that they don’t contain any syntax errors or semantic errors).

Design by Contract

As we saw with the C version of the squareRoot function, programmers can write code to check assertions dynamically. With Eiffel, the compiler can easily identify the assertions because they are separated from the method bodies and it can generate checking code automatically because the compiler understands the assertion syntax. As far as preconditions and postconditions are concerned, the effect of automatic checking is as if we had inserted if statements that check the require part at the start of the method and the ensure part at the end. But what about the invariants? Well, assertions in the invariant clause must always be true for objects of the class. However, in order to check that fact, we need to execute some code, and the only place that code can be executed in an imperative object-oriented language is inside methods. Therefore, invariants need to be checked whenever a method is run. If the invariants are checked at the beginning of each method, we won’t be able to tell if the object itself breaks the invariant; if they’re checked at the end, we won’t be able to tell whether the invariant was already broken when the method was called. Therefore, invariants must be checked at the start of the method (before the preconditions) and also at the end (after the postconditions). The reason for checking before preconditions and after postconditions is that preconditions and postconditions can modify attributes (even though they shouldn’t). In effect, an invariant adds an extra precondition and an extra postcondition to every method. You may be thinking at this point ‘But if we always check that our own methods don’t break the invariants, we can be certain that the invariants are safe when we enter a method’. However, other objects in the system might modify the attributes (by sending them messages), potentially violating our invariants. Since those other objects have no business checking our invariants, we must assume that invariants might have been broken when we start each method.

12.7 DESIGN BY CONTRACT Whenever a piece of software is running, we have to accept that things can go wrong: there may be faults in the program; the user may provide incorrect data; external software that the program depends on may fail; there may be failures in the run-time system, the operating system or the hardware. If we consider that every piece of software is an isolated process running on some platform (operating system plus hardware), we can partition the software universe into activity within the process and activity outside the process. To a large extent, we can control what happens inside the process using good design, good programming guidelines and good testing principles. We can’t, however, control what happens outside the process (even if we access another process that was reliably implemented, the inter-process communication

385

386

Chapter 12

provided by our platform may fail; also, we might depart from our process to access hardware, such as a network or a ﬁle system – things that regularly fail). Before we can hope to write elegant, reliable and efﬁcient software, we must have a clear understanding of the different categories of failure and who is responsible for coping with each category. When we ﬁrst learn to program, most of us assume that nothing will go wrong. Then we run one of our programs, an error occurs and the program falls over, so we learn that it’s better to prevent errors from killing our programs – error recovery – or to make sure that our programs shut down with an explanatory message delivered to the operator – elegant failure. There are many possible failures, so there are many places where we might want to check for errors. To make things worse, best practice encourages us to implement software modules (functions, subsystems, classes) separately, making it difﬁcult to identify who is responsible for detecting errors and who is responsible for handling them. This can result in large quantities of error-checking code, most of which is redundant. bar(39) aClient

aSupplier

Figure 12.1: Sending a message to an object

Let’s look at the worst case in an object-oriented program: design by fear. For example, an object called aClient, inside its foo method, wants to send a message called bar to aSupplier, as shown in Figure 12.1. (Both objects are running in the same process.) Let’s assume that the code for these two objects is written independently (even if the two classes were implemented by the same person, there may have been a signiﬁcant gap between the implementations of each class, long enough for the implementor to have lost faith in the older class). Let’s assume that the implementor of Client is called Beryl and the implementor of Supplier is called Fred. In an ideal world, Beryl would simply write result = aSupplier.bar(anObject) and assume that result was valid. In that same ideal world, Fred would assume that bar was invoked correctly, perform the calculations and return the result. However, realizing that the world is imperfect, both programmers are likely to proceed more cautiously. While writing the foo method, Beryl thinks to herself ‘I’d better check that aSupplier is in an appropriate state to receive the bar message, before I send it’, so she adds code to check the state of aSupplier. Having done that, she thinks ‘I can’t be sure that the parameter I’m about to pass is valid, so I’d better check that too.’ Having coded two checks, she adds the code to send the message. For the returned result, Beryl thinks ‘I can’t be sure that the result I get back is valid’, so she adds code to test the result. However, Beryl is being thorough. She decides that she can’t trust aSupplier to be in a valid state after replying to the message – after

Design by Contract

all, the message may have had an undesired effect that wasn’t predicted by its implementor. So, Beryl adds a fourth chunk of error-checking code to ensure that aSupplier is valid before proceeding. Beryl’s foo method would look like this in Java: void foo() { ... if (! ... check aSupplier ... ) { fail("Invalid state of a supplier"); } if (! ... check anObject ... ) { fail("Invalid parameter for a supplier message"); } int result = aSupplier.bar(anObject); if (! ... check result ... ) { fail("Invalid result from a supplier message"); } if (! ... check aSupplier ... ) { fail("Invalid state of a supplier"); } ... }

In order to avoid nested if statements, the logic of the checks has been ﬂipped using Java’s ‘not’ operator, !. The fail function is a method on Client that prints the class name, the name of the failing method and the parameter passed in and then shuts down the process with System.exit(-1). When Fred writes the bar method, he also does his best to write robust, fail-safe code. Before he performs any calculations, he checks that the current object is in a valid state to receive bar. Next, he checks that the parameter is valid – after all, it would be foolish to trust the client. Once he is sure that the receiver is ready and that valid data has been passed, Fred adds code to calculate the result. Before the return statement, Fred decides that it would be a good idea to check that the result is valid – because there may be faults in the calculations that he’s written or in other code that he’s called. Still not satisﬁed, Fred adds one more piece of checking code, to make sure that the calculations haven’t corrupted the current object. Fred’s bar method would look like this: public int bar(Object anObject) { if (! ... check this object ...) { fail("Invalid state of this object"); } if (! ... check anObject ...) {

387

388

Chapter 12

fail("Invalid parameter from a client"); } int result = ... if (! ... check result ...) { fail("Invalid result for a client"); } if (! ... check this object ...) { fail("Invalid state of this object"); } return result; }

The end result is a lot of code – perhaps more error-checking code than normal code: the state of aSupplier is checked twice before the calculation and twice after; anObject is checked twice; result is checked twice. Although you may not be quite as thorough (or paranoid) as Fred and Beryl, their code may still be uncomfortably familiar. This situation applies equally well to foo and bar written as stand-alone functions, procedures or subroutines. Error-checking code can become even more verbose if we employ any of the common techniques listed below: • Result codes: A result code is a number returned from a routine to indicate success or failure – for example, zero indicates success, a negative number indicates failure. Where used, result codes are typically applied to every kind of routine, regardless of whether it would normally return a result. This leads to many ‘If result code is okay’ checks. • Global error variables: Here, rather than returning a success value, programmers set a global value instead (errno in the C function library is a good example). Functions return their values as normal, setting the error variable if there is a problem. However, this leads to multiple ‘If the error variable is non-zero’ checks. (An object-oriented equivalent of a global error variable would be a class attribute on a class called Error.) • Exceptions: An exception is an error signaled by supplier code that causes the program to jump to handler code that is kept separate from the normal path code. Used well, exceptions can remove error-checking clutter from normal code. However, without a design rule that speciﬁes who is responsible for handling what, exceptions make matters worse. In the 1980s, Bertrand Meyer described how to combine the best features of formal methods and object-oriented programming in a pragmatic way, that makes it easy to produce code that is robust, elegant and efﬁcient. Meyer called his technique Design by Contract.

Design by Contract

Although Meyer described Design by Contract in the context of Eiffel, implemented using assertions and dynamic checking, we can use the same techniques in other languages – we’ve already seen some of them being used in C. (Don’t be put off by the use of the word ‘design’ – the idea of a contract is relevant to all phases of software development.) With Design by Contract, we’re asked to imagine a binding contract between a client object and a supplier object, with obligations on both sides. The deal is that, as long as the client fulﬁlls its obligations, the supplier will too (see Figure 12.2). bar(39) aClient

aSupplier

Figure 12.2: A contract between client and supplier

Using formal speciﬁcation terminology, we can be precise about contractual obligations: • • • •

The client must respect the supplier’s invariants. The use of a supplier method is subject to its preconditions. The implementation of a supplier method must guarantee its postconditions. The implementation of a supplier method must guarantee the supplier’s invariants.

While a method is doing its work, it is allowed to break the contract, as long as the contract is fulﬁlled when the work is ﬁnished. For example, if we have a Supplier invariant that a + b = 4 and both attributes have the value 2 when bar is called, it would be valid for bar to return with a set to 10 and b set to -6; however, since we can’t set both values at the same time in most programming languages, a + b will spend some time inside bar as 12 or -4. A class can also have an internal contract, in the form of invariants for nonpublic attributes and preconditions and postconditions for nonpublic methods. An internal contract is for the beneﬁt of implementors and maintainers; it is much less important than the external contract.

12.7.1 Contracts and Inheritance Since object-oriented programming supports inheritance, we should consider the effect of inheritance on contracts. Because of polymorphism, we must guarantee that each subclass

389

390

Chapter 12

contract is the same or better, from the client’s point of view, than that of its superclass (or superclasses, in the case of multiple inheritance). Without this rule, we could spring a nasty surprise on a programmer that was using a subclass object via a superclass variable: for example, if the contract for class Customer speciﬁes that all customers have at least two bank accounts and we introduce a SimpleCustomer subclass with only one bank account, anyone accessing a SimpleCustomer object via a Customer variable might try to access a second, nonexistent, bank account. As we move down the class hierarchy, class invariants are combined using ‘and’: we can only add to the invariants. For example, if class X has an invariant i1 and subclass Y adds an invariant i2, the total invariant on Y is i1 and i2. The implication is that we’re providing extra guarantees about the attributes. Preconditions and postconditions are only an issue if we redeﬁne the associated method in a subclass. For preconditions, we must ensure that the redeﬁned method accepts, at least, all of the messages that it accepted in the superclass: we can only weaken the preconditions. Therefore, any preconditions that we add to a redeﬁned method are combined with those on the inherited method using ‘or’. For example, if foo has a precondition pre1 and a subclass adds precondition pre2, the net effect is that the redeﬁned method has the precondition pre1 or pre2. For postconditions, we must ensure that each subclass guarantees at least as much as the superclass: we can only strengthen the postconditions. Therefore, postconditions are combined using ‘and’. For example, if foo has a postcondition post1 and a subclass adds post2, the redeﬁned method will have a combined postcondition of post1 and post2. • If invariants are omitted, it means that ‘nothing is guaranteed for this class’; if we add invariants to a subclass when there are none on the superclass, the net effect is the invariants that we added to the subclass. • If preconditions are omitted, it means that ‘this message applies in all cases’; if we add preconditions to a redeﬁned method that previously had none, the new preconditions will have no effect (we can’t weaken a method that already accepts everything). • If postconditions are omitted, it means that ‘this message doesn’t guarantee anything’; if we add postconditions to a redeﬁned method that had none, the net effect is the postconditions on the redeﬁned method (anything that we add strengthens what was previously an empty guarantee). With the help of Design by Contract ideas and formal speciﬁcation terminology, Beryl, our paranoid programmer, can make her code much clearer: void foo() { ... if ( ! ... check aSupplier’s invariants ... ) {

Design by Contract

fail("Invariants broken for a supplier"); } if (! ... check bar’s preconditions ... ) { fail("Preconditions broken for a supplier message"); } int result = aSupplier.bar(anObject); if (! ... check bar’s postconditions ... ) { fail("Postconditions broken by a supplier method"); } if (! ... check aSupplier’s invariants ... ) { fail("Invariants broken by a supplier method"); } ... }

Similarly, Fred can change his code to look like this: public int bar(Object anObject) { if (! ... check invariants ...) { fail("Invariants broken"); } if (! ... check preconditions ...) { fail("Preconditions broken"); } int result = ... if (! ... check postconditions ...) { fail("Postconditions broken"); } if (! ... check invariants ...) { fail("Invariants broken"); } return result; }

12.7.2 Reducing Error-Checking Code The main advantage of Design by Contract is the clear separation between the responsibilities of the client and the responsibilities of the supplier. Before the bar message is sent, the client is in control. Thus the client has responsibility for ensuring that the contract is not broken before the message is sent. While the method is executing, the supplier is in control. Thus the supplier must ensure that the contract is not broken by the time the method returns.

391

392

Chapter 12

Let’s apply this idea of responsibility to a more optimistic Fred and Beryl. This time Fred and Beryl trust each other: Fred writes the bar method assuming that Beryl won’t break the contract; Beryl, for her part, uses the result of bar directly, conﬁdent that Fred won’t break the contract either. Beryl’s code can now look like this: void foo() { ... if ( ! ... check aSupplier’s invariants ... ) { fail("Invariants broken for a supplier"); } if (! ... check bar’s preconditions ... ) { fail("Preconditions broken for a supplier message"); } int result = aSupplier.bar(anObject); ... }

And Fred’s code can look like this: public int bar(Object anObject) { int result = ... if (! ... check postconditions ...) { fail("Postconditions broken"); } if (! ... check invariants ...) { fail("Invariants broken"); } return result; }

This is excellent news: Design by Contract has shown us how to cut our checking code in half, by removing all redundancy. This results in faster development, better performance, fewer faults and easier maintenance. Can we use Design by Contract to reduce checking code even further? Yes. There are many occasions when a piece of code that we write can’t possibly break the contract. For example, consider the following implementation of Supplier: // A supplier // Invariant: x < 5 public class Supplier { private int fieldX; /* Constructor */

Design by Contract

public X() { fieldX = 1; } /* Getter for x Preconditions: true (call this method any time) Postconditions: true (no side effects) */ public int getX() { return fieldX; } /* A method Preconditions: o != null Postconditions: For result r, r > x */ public int bar(Object o) { fieldX = o.toString().length() % 5; return fieldX + 1; } }

In this example, x refers to a UML-style attribute – the value of x is accessible to the outside world via the message getX. In order to emphasize the difference between the public attribute and the private implementation, the ﬁeld has been given a different name, ﬁeldX. Because of the way it’s been written, the Supplier code always respects invariants and postconditions: • After the creation of a Supplier, x is less than 5 because ﬁeldX is set to 1 by the constructor. • After the execution of getX, x is less than 5 because getX has no side effects. • After the execution of bar, x is less than 5 because ﬁeldX is set to the remainder after dividing something by 5. • After the execution of bar, the result is greater than x because bar returns ﬁeldX + 1. Since Supplier can’t possibly break its half of the contract, no checking code is needed. (In principle, platform failures could break the contract, but it would not be feasible for us to add checking code for such situations.) Now, consider the following implementation of foo, on Client: void foo() { Supplier aSupplier = new Supplier(); int result = aSupplier.bar(new Plate("Wedgwood")); ... }

When the bar message is sent, we know that foo can’t have violated aSupplier’s invariants, because it hasn’t touched the new object’s attributes. Also, we know that foo can’t have

393

394

Chapter 12

violated bar’s preconditions, because it’s passing in a non-null object. Thus, we don’t need any checking code in foo either. The end result is a program with no checking code at all. We’ve now seen that well-deﬁned obligations, mutual trust and the nature of the code that we’re writing can halve the amount of error-checking code, or even remove it altogether. What we haven’t covered yet is how we should deal with the fact that our software exists in an imperfect world: client and supplier code can contain faults; if we jump out of our process to another process or another piece of hardware, nasty things can happen that are beyond our control. Faults within our process can be dealt with by enforcing the contract; problems outside our process can be dealt with by constructing application ﬁrewalls. We’ll look at each of these techniques in turn.

12.7.3 Enforcing the Contract Design by Contract relies on the assumption that as long as the client object meets its obligations, the supplier will meet its obligations too. But programmers are fallible; sometimes obligations won’t be met, so we need dynamic checking to detect faults. But who is responsible for checking that contractual obligations have been met? And, what should we do if we detect a failure? In the client and supplier example, the bar method should be responsible for checking postconditions and invariants before it returns, because only the bar method or one of its suppliers could have broken the contract between the start of the method and the end. Therefore, dynamic checking for the supplier half of the contract (postconditions and invariants) should appear at the end of the bar method. The question of who should check the invariants and the preconditions before bar starts is more complicated. Client code is responsible for making sure that the client half of the contract is not broken. But, if we conclude that every client of bar has to have contract-checking code, the code will be duplicated all over the system. Object-oriented theory states that responsibilities should be assigned to the relevant object. This suggests that checking the client side of the contract should be a service provided by the supplier (because the preconditions and invariants are associated with the supplier). Thus, in order to avoid duplicating the client’s contract-checking code, we should place it at the start of bar. You can view this outcome in one of two ways: either bar is protecting the supplier from faults in the client or bar is providing a service for the client, to check that it has no faults. (In terms of Design by Contract, the second interpretation is preferable.) Combining informal methods, Design by Contract, trust and dynamic checking, a Java implementation of the bar method looks like this: public int bar(Object anObject) { if (! ... check invariants ...) { fail("Invariants broken");

Design by Contract

} if (! ... check preconditions ...) { fail("Preconditions broken"); } int result = ... if (! ... check postconditions ...) { fail("Postconditions broken"); } if (! ... check invariants ...) { fail("Invariants broken"); } return result; }

With the implementation above, no contract-checking code is needed in the client.

12.7.4 Application Firewalls As long as we stay within one process, we can rely on contracts to provide elegance, robustness and performance. When we jump out of our process, however, there is no contract enforcement to protect us. Therefore, we should be paranoid about everything we access beyond our process boundary, including, at least: • • • • • • •

other processes the operating system user interfaces ﬁle systems networks databases devices

We can construct an application ﬁrewall at the boundary to protect the code inside our process, for example, a user interface ﬁrewall would check the validity of user data before passing it on to business objects, while a database ﬁrewall would catch database errors. (These have nothing to do with Internet ﬁrewalls, which protect an intranet.) Sometimes, we can prevent invalid data crossing an application ﬁrewall in the ﬁrst place. For example, if a user must not try to log in until they’ve typed in a user name and a password, we can disable the Login button in our GUI until they have done so.

395

396

Chapter 12

12.8 INFORMAL SPECIFICATION IN JAVA If we’re implementing our designs using a language such as Eiffel, best practice (using informal speciﬁcations and dynamic checking) is built into the language itself. When using other languages, we should apply similar principles, but we have to write the dynamicchecking code by hand. In this section, we’ll see how to apply speciﬁcation theory to Java, since Java is a common, pragmatic alternative to Eiffel.

12.8.1 Documenting a Contract using Comments We should always, at the very least, document the contract of our classes using comments in the source code. This is because our code will end up being used, reused, or maintained by programmers who look at the source code rather than anything else. Any contract-related comments in the source code should also appear in the design and speciﬁcation artifacts. Every class should have a lengthy comment at the top – the class comment – that describes what the class is for and any invariants for its public attributes. Be aware that other programmers looking at the source code may not be familiar with formal jargon, so you might want to use phrases such as For all objects of this class, the following conditions apply:. The class comment can also contain examples of how the class should be used. Every public message on the class should have a comment at the top – the message comment – that describes what the corresponding method is for and lists any preconditions and postconditions. For the sake of uninformed readers, you might prefer to use terms like requires and guarantees rather than preconditions and postconditions. Always keep in mind, when implementing a method, that it’s the caller’s responsibility to make sure the invariants and preconditions are met before the method is called. In other words, your coding should be optimistic. The ﬂip side of this, of course, is that you should make sure that your implementation doesn’t violate the postconditions or the invariants.

12.8.2 Checking Conditions Dynamically In principle, it’s a good idea to add an invariant check and a precondition check to the start of every public method, and a postcondition check and an invariant check to the end. However, in the absence of special language mechanisms, such as those provided by Eiffel, you may prefer to be more selective. It’s a good idea to add checks to the start of any public method that you expect to be reused heavily (if the code is destined for a library, for example). Third parties will appreciate the extra help that your checking code provides, as they become familiar with your classes. As for checks at the end of public methods, this is a matter of personal preference. If your method couldn’t possibly break the supplier half of the contract, the checks would be redundant. If you think that your code could break the contract, perhaps you don’t

Informal Speciﬁcation in Java

trust yourself: as your experience with implementing object-oriented code increases, your self-belief should grow, so such checks should also become redundant. The only compelling reason to include end-of-method checks is if you believe that thirdparty code invoked by your method could break the invariants, by corrupting attributes: since most code is invoked in a client–server style, this should not happen often.

12.8.3 Signaling Contract Violations using RuntimeExceptions Whatever choices we make about enforcing the contract, we need a clean, elegant way of signaling contract violations. Thus far, the issue has been side-stepped through the use of a fail message. But what code do we need inside the corresponding fail method in order to fail elegantly? One possibility is to quit the program: in Java we can do this with System.exit(-1). However, this technique discards most of the information about how we arrived at the failure point – we might be told which method was at the end, but not which method called that, or which method called that, and so on. Instead of exiting explicitly, a Java method can throw an exception, a signal to the calling method that the current method can’t proceed. Java has a particular type of exception, called a RuntimeException, that we can use in such cases. A RuntimeException should be used to signal that the method can’t proceed because of a fault in its implementation or in the calling method. The calling method is not expected to catch the RuntimeException: instead, it should allow the run-time system to re-throw the exception to its own calling method, where it will be re-thrown to the method that called that, and so on; eventually, the exception will reach the skin of the process, the main method, having halted each method on its way out. When a RuntimeException is thrown by main, the program will be stopped by the run-time system and the user (or tester, or administrator) will be shown a trace through the failing methods, along the lines of: RuntimeException in bar() "Preconditions violated" main() called foo() on Client foo() called bar() on Supplier

This is exactly the kind of information we need to start debugging. The Java style for signaling a broken contract is shown in the following example: public class Supplier { ... // Precondition: o != null. public int bar(Object o) { if (o == null) { throw new RuntimeException("Preconditions violated"); }

397

398

Chapter 12 ... } ... }

(You can throw a subclass of RuntimeException, such as IllegalArgumentException, if you want to provide more detail to the user.) When Java code is running inside a larger program, such as a servlet running inside a Web server or a business component running inside a GUI, we shouldn’t allow RuntimeException objects (or any other kind of exception), to reach the main method, otherwise the whole program will shut down. Instead, generic code at the edge of the program can catch unhandled exceptions and report them to the administrator or the user, or both. For example, a Web server can append an error report to the system log ﬁle and then display an error message to the user, along the lines of Your request can’t be completed because ..., plus some advice on what to do next.

12.8.4 External Systems As we’ve seen, while running code within our own process, we can safely disable dynamic checking. Trust your own code; trust your colleagues’ code; trust the code in patterns, libraries and frameworks. However, we still need to construct application ﬁrewalls to protect our code from the outside world. Application ﬁrewall code will never be switched off, even after testing. External systems can be categorized as clients or suppliers – we’ll look at how to deal with each of these in turn.

External Clients Requests coming from external clients are destined for our business objects, so the contracts of our business objects must be respected. Thus, the application ﬁrewall should check the client half of the contract, on behalf of the business objects. However, the knowledge of the contract is best encapsulated in the business objects themselves (because that’s where the knowledge should reside). We can resolve this apparent paradox by encapsulating the invariant and precondition checks in messages on the business objects. For example, consider the following class: // Invariant: i1 public class Foo { public boolean invariantOK() { ... // Return true if i1 is okay } public boolean okForBar(s, f) { ... // Return true if s, f satisfy p1 }

Informal Speciﬁcation in Java

// Precondition: p1 public void bar(String s, float f) { if (!invariantOK()) { throw new RuntimeException("Invariant violated"); } if (!okForBar(s, f)) { throw new RuntimeException("Precondition violated"); } ... } }

Code in an application ﬁrewall can extract the values for bar’s parameters (from a user interface or a servlet, for example) and use the service messages on Foo to test whether invoking bar would be valid, as in: // Application firewall code if (aFoo.invariantsOK() && (aFoo.okForBar(s, f)) { aFoo.bar(s, f); } else { ... // Signal error to client }

As a variation on this idea, you could provide checking and non-checking versions of each ordinary Foo message, for example, bar and barWithChecks. This would move the application ﬁrewall code into Foo, as a service for clients. The downside of this approach is that you would be providing the checking versions of the messages even though they might never be used – they would only be used if the object ended up being invoked from an application ﬁrewall. Another variation would be to put the checking versions of the messages into a separate class, called FooWithChecks perhaps. This approach is better than code in the application ﬁrewall itself but you would be providing an entire class that might never be used.

External Suppliers When we’re jumping out of our process to access an external supplier, we must check the state of the external supplier and we must check any information that we get back. Again, the notion of an application ﬁrewall is useful here. But, what would we do if the application ﬁrewall detected an error? We will have been invoked originally by some client code and, if the application ﬁrewall detects an error, the client code must be told that there is a problem. A good way to do this, avoiding the messiness of result codes or error variables, is to throw an exception. However, unlike the ‘broken contract’ situation, we don’t want client code to ignore this kind of exception – external problems are a predictable fact of life, even in fault-free code. Therefore, we must force the client to handle such cases.

399

400

Chapter 12

To help with this, Java has a second kind of exception – represented by the Exception class – that the compiler will force client code to handle. In Java, Exception objects are called checked exceptions because the compiler requires checking code to be provided, while RuntimeException objects are called unchecked exceptions because no checking code needs to be (or should be) provided. So, in order to signal a failure from an out-of-process supplier, we can do something like the following: public class ExternalResourceUser { public void useResource() throws Exception // Checked exceptions must be listed { ... // Use the external resource ... // Detect a problem throw new Exception("External resource failure"); } }

As with RuntimeException, you can use subclasses of Exception, such as IOException, if you want to be more speciﬁc.

12.8.5 Enabling and Disabling Dynamic Checks Dynamic checking has implications for the performance of your system: checking preconditions, postconditions and invariants in every method could easily make the code ten times slower than it would otherwise be. We aim to produce software that, when deployed, will never violate an assertion – if no assertion is ever violated, there’s no need to do any checking. During development, on the other hand, we expect violations to happen often. What we need, in order to satisfy the different requirements of development and deployment, is a way to turn dynamic checking on and off. Eiffel implementations vary in whether they use a run-time switch or a compile-time switch, but it is relatively simple to be selective about what we turn off and what we leave on. For example, we may turn postconditions off for the library code that we’re using during development (because we expect the library to have been debugged already). With non-Eiffel languages, we have to work harder. There are two ways to enable and disable dynamic checks. The ﬁrst involves ﬂipping a run-time switch (using a command-line parameter or an environment variable): if the switch is on, the run-time system will do the checks; if the switch is off, the run-time system will skip over the checks. The second method uses a compiler switch to build two versions of the system, one with checking code included and one without. In both cases, we can be more selective: for example, we might disable all postcondition checks or all the checks in library code.

Informal Speciﬁcation in Java

In principle, Java implementations could use either a run-time or a compile-time switch but, in practice, the situation is a little more complicated. Java programmers can only use a compile-time switch by careful programming; a run-time switch (the assertion mechanism) is available, but it shouldn’t be used to check all parts of the contract.

Implementing a Compiler Switch We can simulate a compiler switch using a well-documented trick. First, we introduce a class that stores the value of the switch as a class constant (a class ﬁeld with a ﬁxed value). This is how such a class might look (where ﬁnal means ‘constant’): public class ContractSwitch { public static final DO_CHECKS = true; // It’s okay to have a public field if it’s final }

Now, any dynamic checking code can be wrapped in an if statement that depends on the value of the switch, as in: public class Supplier { ... // Precondition: o != null. public int bar(Object o) { if (ContractSwitch.DO_CHECKS) { if (o == null) { throw new RuntimeException("Preconditions violated"); } } ... }

When the compiler gets to the if (ContractSwitch.DO CHECKS) statement, it spots that the class constant is true, so it knows that the body of the if statement will always be executed: the precondition check is compiled and included in the current build. However, no code needs to be generated for the outer if statement because it would be redundant at run time. The net effect is that the current build includes the contract-checking code but not the code which checks the switch. In order to build a version of the system with checks disabled, we ﬁnd the deﬁnition of the ContractSwitch class, set DO CHECKS to false and recompile our system. Now, when the compiler encounters the if (ContractSwitch.DO CHECKS) statement, it will conclude that, since the constant is false, the outer if statement will always fail, so there is no point generating code for it or anything inside it. The net effect is a build of the system that contains no checking code and no switch-checking code. This compiler-switch technique is part of the Java language speciﬁcation, so we can expect it to work with every Java compiler. With a little thought, you should be able to work

401

402

Chapter 12

out how to make the technique more selective (by adding more class constants). Because Java compilers compile classes by package, we must add a ContractSwitch class to every package.

Using the Assertion Mechanism Java has an assertion mechanism that allows the programmer to insert statements such as assert o != null. A run-time switch can then be used to control whether assertions are on or off, i.e. whether assertion failure should cause the program to stop automatically. Although a run-time switch is used, it is possible for Java implementations, when the switch is off, to remove all the process size and performance overheads normally associated with assertions. Traditionally, assertion mechanisms in programming languages have been used to enforce contracts and to insert fault-checking code at arbitrary points. According to the documentation for Java’s assertion mechanism, it should not be used to enforce the client half of a contract. The reason given is that some checks are too important to allow them to be disabled (for reasons of safety). A better recommendation would have been to state that some checks, which reside in application ﬁrewalls, are mandatory and the assertion mechanism should not be used for those. Such a recommendation would have been compatible with everything else that we’ve seen in this chapter. Despite the fact that the assertion mechanism has all the characteristics of a switchable mechanism for checking contracts, we have to accept that we shouldn’t use it for this. Other programmers who have read the documentation won’t expect it to be used in this way. Also, if assertions are on when a check fails at run time, the kind of exception that is thrown is an Error not a RuntimeException. An Error is an unrecoverable problem that programmers should not try to handle, such as a fault in the run-time system or running out of memory. As the assertion mechanism throws an Error, it doesn’t ﬁt in with the guideline that application faults should be signaled with RuntimeException objects. This is a shame. Although automatic contract enforcement may be added to Java in the future, for now, we’re stuck with the compiler-switch technique.

12.9 SUMMARY In this chapter, we’ve looked at: • Speciﬁcations, which are complete, unambiguous descriptions of the required behavior of our software. Formal speciﬁcations are scientiﬁc and rigorous and use specialized language; informal speciﬁcations are pragmatic and partial, and may be expressed in natural language or in the syntax of a language such as Eiffel.

Review Questions

• How speciﬁcations can be written in an object-oriented manner. An object-oriented speciﬁcation allows us to assert the preconditions, postconditions and invariants for a class. • Design by Contract, which combines formal methods and object-oriented programming in a way that makes it easy to produce robust, elegant and efﬁcient code. A binding contract between a client object and a supplier object imposes obligations on both sides. • How to use speciﬁcations and Design by Contract in Java, particularly by using comments and Exception objects. We also considered the Java assertion mechanism and discarded that as a potential solution.

FURTHER READING For a larger helping of OCL, see [Clark and Warmer 02] and the OCL speciﬁcation itself [OMG 03b]. For a description of Design by Contract from the inventor himself, see [Meyer 97]. Bertrand Meyer also covers every aspect of object-oriented programming. For a good discussion of what to do and what not to do in Java, including the proper use of exceptions, see [Bloch 01].

REVIEW QUESTIONS 1. What is a ‘class invariant’? Choose only one option. (a) A class whose source code is versioned and therefore cannot be changed. (b) A class whose objects have constant ﬁelds. (c) A condition that will always be true for an instance of the class. 2. What is meant by the term ‘design by fear’? Choose only one option. (a) Design is scary. (b) You cannot know when to trust the code. (c) You design a system too quickly because of time pressures. 3. What is ‘Design by Contract’? Choose only one option. (a) Designing code as if there were a contract between an object that sends a message and the object that receives it.

403

404

Chapter 12

(b) Reinforcing the contract between every pair of objects by increasing the amount of error-checking. (c) Protecting your software using a contract with a ﬁrewall. (d) Designing a software system under contract.

ANSWERS TO REVIEW QUESTIONS 1. A ‘class invariant’ is c. A condition that will always be true for an instance of the class. 2. The term ‘design by fear’ means that b. you cannot know when to trust the code. 3. ‘Design by contract’ means a. Designing code as if there were a contract between an object that sends a message and the object that receives it.

13 Continuous Testing

I

n this chapter, we’ll be covering all aspects of testing. As with other areas of software development, the discipline of engineering is often used as a guide.

Learning Objectives Understand the many terms used in testing.

•

Consider a testing strategy.

Work through an example of test-driven development.

Chapter Outline 13.1 Introduction 13.2 Testing Terminology 13.3 Types of Test 13.4 Automating Tests 13.5 Preparing for Testing 13.6 Testing Strategies 13.7 What to Test For 13.8 Test-Driven Development 13.9 An Example of Test-Driven Development using JUnit 13.10 Summary Further Reading

13

408

Chapter 13

13.1 INTRODUCTION Developing software is a complex business. No matter how hard we try, we won’t be able to eliminate all faults (or bugs) simply by going through the phases of requirements, analysis, design, speciﬁcation and implementation. However, through good practice, we can make sure that the most serious faults do not occur in the ﬁrst place. In addition, we need a separate testing phase, with the goal of eliminating all remaining faults before release. Nowadays, it is generally accepted that tacking a testing phase on to the end of the development is inefﬁcient: we must also test our code, and other artifacts, as we go along. We should ensure that many different people are involved in testing: developers; peers (colleagues not directly involved in the current project); customers; project managers; testers (colleagues with primary responsibility for the testing phase). Essentially, there are three phases of testing activity: during development (by developers), during the testing phase (by an expert testing team) and after release, when all users and developers gather feedback and ﬁx faults discovered while the software is live. Since this book is aimed primarily at members of the development team, we will look at test-driven development, a form of continuous testing where developers test their code as they go along. The advantages of this approach are that: • It improves the quality of the software (the more testers the better). • It reduces the cost of the testing phase. • It shows the programmers that they’re making real progress (rather than just producing lines of code). • It reduces the number of (embarrassing) faults that are linked to the programmer during the testing phase. • It helps programmers to refactor (reorganize) their code, for style or performance reasons, without breaking anything that they’ve already written.

13.2 TESTING TERMINOLOGY • A test checks that some aspect of our software is correct, such as ‘Test whether we can log on’, ‘Test the purchasing subsystem’ or ‘Test that the memory footprint doesn’t exceed 500 MB under maximum load’. • An error is a mistake made by a programmer, based on a misunderstanding, for example. An error usually gives rise to one or more faults. (This is distinct from using ‘error’ to

Testing Terminology

•

• • • •

•

mean an undesirable condition, expected or otherwise, that occurs while the system is running – see the deﬁnition of failure.) A fault is an incorrect piece of code: in the same way that a fault in a car’s engine stops it running smoothly, a software fault stops the system from operating correctly. For example, a method written on the assumption that arrays are indexed from 1 has a fault if arrays are actually indexed from 0. The terms fault or defect are used in preference to bug, which is considered to be rather informal. Faults of omission occur when the developer has left something out (such as a customer requirement or code that should deal with a particular situation); faults of commission occur when there is a problem committing requirements to code, design or speciﬁcation (such as stating that the area of a triangle is base times height or attempting to read from a ﬁle that has been closed). A failure is a system malfunction, usually caused by one or more faults, for example, ‘Someone has hacked into the system because of the buffer overrun fault’. A ﬁx is when a fault is repaired. Veriﬁcation checks that the software is correct with respect to the requirements documentation (system use cases). Validation checks that the software is what the customer requires, i.e. that it performs the functions that the customer needs in a way that is acceptable to the customer and to the end users. A speciﬁcation is a description of the interface of a piece of code, which lays out what can go in and what must come out. Therefore, speciﬁcation testing is an umbrella term for black-box testing and use case testing.

13.2.1 Black-Box Testing With black-box testing, whatever is being tested (the system, subsystem, class, or method) is treated as an impenetrable object (see Figure 13.1). The only access to the box is via its published interface – messages, parameters and so forth. Therefore, all the tester can do to assess the correctness of what the box does is to capture and analyze the replies (from messages); examine any external side effects (such as the creation of entries in a database); and check the time taken to answer requests or execute commands. Although each use of the interface may cause internal side effects (as the box changes from one state to another), the tester can only view the effect that these internal side effects have on the subsequent use of the interface: even if details of the internals of the box are available, they’re ignored for the purposes of black-box testing. Black-box testing arises from the philosophy of ‘we don’t care how the code achieves its ends, as long as it achieves them.’ This ﬁts well with the idea that the system requirements that are written down before the software is produced are the most important aspect of software

409

410

Chapter 13

Inputs

Outputs

Side-Effects

Figure 13.1: Black-box testing

development: beyond the requirements phase, the analysts, designers and implementors have free reign to fulﬁll the requirements in any way they see ﬁt. Black-box testing insulates the tester from the internal complexity of the software. A disadvantage of black-box testing is that, because the internal structure and lines of code are ignored, potential improvements may be missed – after all, it is possible for a piece of software to be correct and adequate without being good quality. Another disadvantage is that faults that are not caught by the tests themselves will remain undiscovered until much later. Other terms for black-box testing include behavioral testing, emphasizing that the testing concentrates on how the box behaves, from an external point of view, rather than on how that behavior is achieved (which would require knowledge of the inside of the box), and functional testing.

13.2.2 White-Box Testing In white-box testing, we’re allowed to look inside the box, examining structure and detail right down to individual lines of code (see Figure 13.2). The beneﬁts of white-box testing include the fact that peers (other software developers) can propose improvements, such as refactoring, for better performance; easier maintenance; and more opportunities for reuse. Another beneﬁt is the fact that deﬁciencies may be spotted early or spotted when they might otherwise be missed (because it is not possible

Types of Test

Inputs

Outputs

Side-Effects

Figure 13.2: White-box testing

to test software exhaustively). Deﬁciencies spotted by white-box testing include classic programmer mistakes, such as traversing a collection of values more times than is necessary, and incomplete or incorrect interpretation of the requirements. Since, during the course of development, the structure and content of code will change, white-box testing is best left until late in the development, usually during the testing phase. This is particularly true of development that is spiral, iterative and incremental, where we accept that analysis, design, coding and system requirements will need to be adjusted as we learn more about the problem and its solution. Other terms for white-box testing include structural testing, reﬂecting the fact that the internal structure of the box is considered as well as its external behavior, and glass-box testing.

13.3 TYPES OF TEST Most testing experts break the testing process up into a hierarchy of distinct levels: • Unit testing is the lowest level of the hierarchy. A unit is a single, coherent piece of code (traditionally, a function or a procedure; in object-oriented software development, a class). • Integration testing checks how independent pieces of code work together.

411

412

Chapter 13

• System testing checks the operation of all the subsystems working together, along with any interaction they have with their environment. The most important form of system testing is use case testing because the system use cases describe completely what the system must be able to do. Having designed tests for each use case, we can go on to provide tests that check whether the use cases work in parallel.

13.3.1 Unit Testing For each class, we design tests that check whether it works correctly, efﬁciently and elegantly. While designing unit tests, we aim to check that instances of the class have the correct message signatures (name, return type, parameter types and parameter names) and that they generate the correct replies from our test data. We also want to check that each object starts in a valid state and moves from one valid state to another, as it is used. (Recall that the state of an object is recorded in the values of its attributes.) For black-box testing, we can only verify the state using messages (getCustomerId, for example); if we do white-box testing, we can check the ﬁelds and we can look inside the methods for faults, poor style or poor quality. If the class makes use of class ﬁelds and class methods, we would have to check the class itself as well as its objects. With unit testing we have a dilemma: how do we test the unit without testing its collaborators too? In object-oriented terms, we may decide that encapsulated objects are simply part of the unit and therefore take no special steps for black-box testing. (For whitebox testing, we can simply ignore the collaborators.) However, if we wish to be pedantic, we can replace the collaborators with simulated, or mock, objects that provide the bare minimum behavior needed to test the original class. We may also need to use mock objects (or stubs, as they’re sometimes called) as stand-ins for code that is not available yet (because it hasn’t been implemented or because it requires some conﬁguration, such as a database, that we don’t want to be concerned with for the moment).

13.3.2 Integration Testing In a ﬁnished system, every object collaborates with other objects, so integration testing is a necessary step after unit testing (especially since different classes are generally written by different developers). For example, we might want to test how well the objects in our business layer work together. Integration testing usually takes place at one or more of the following levels: • • • •

collaborating classes that form a logical unit layer package library

Types of Test

• framework • system or subsystem Subsystem testing is a form of integration testing where the collection of classes under test comprise a whole subsystem (as deﬁned by the system designers).

13.3.3 Alpha Testing With alpha testing, we accept that the system is a work in progress. The aim is to make sure that we’re developing something that is likely to be correct and useful to the customer. For example, we may implement some core functionality within the framework of our system design and then make it available to colleagues and carefully-selected customers to try out. If the feedback is good, we know that we’re on the right track (and we will gather useful information about improvements); if the feedback is particularly negative, we must go back to the drawing board, otherwise we will end up delivering a system that is, at best, useless or, at worst, something that the customer will refuse to pay for. Alpha testing is often performed against a prototype, i.e. a version of the system that is produced quickly and cheaply. Thus, if the alpha testing fails, we have not wasted too much time or money. Once the alpha testing is complete, we should discard the prototype code in favor of the more elegant solution that we will produce during the full-scale development. Although it can be hard to throw code away, we must remind ourselves that the prototype was meant to be developed quickly, just for experimentation – it is highly unlikely, then, that the design or coding will be top quality. Although we throw the code away, we retain everything that we have learned about the problem and its solution up to that point (draft use cases and UML artifacts, for example).

13.3.4 Beta Testing Beta testing occurs when an entire increment of the system is complete, after all phases of development and in-house testing. It involves asking people to try out the software in its real environment. We do this because we want to ﬁnd the most obvious faults and ﬁx them, before ﬁnal release. No matter how good our development process is, there will usually be faults that have not been discovered by the developers or the special-purpose testing team. Normally, faults are missed because exhaustive testing is impossible (or at least infeasible), which means that testers must use their skill and judgment to anticipate the way that the software will be used. Beta testing is used to ﬁll in the gaps between anticipated use and actual use. Beta testing should be carried out by end users, not by any member of the testing team. Depending on the nature of the software, beta testers may be colleagues, real customers, or both. Since beta testing and producing feedback can be frustrating tasks, we normally need

413

414

Chapter 13

to offer some kind of incentive. For colleagues, we might try ‘If you beta test my software, I’ll beta test yours’ or simply ‘It’s for the good of the company’; for real customers, we might offer a discount on the released version. For a system with only one eventual customer, we may have to persuade them that beta testing is a normal part of the process and doesn’t imply poor quality – the beta version could then be tested by a few of the customer’s employees before wider release within their organization.

13.3.5 Use Case Testing With use case testing, we take each use case in turn and design tests that conﬁrm that the system satisﬁes the use case. We may do this by considering a number of possible scenarios for each use case, each scenario being a legal sequence of events – hence the term scenario testing. (Although a use case may appear to be a single sequence of events, it will often describe alternative paths and repetitions – therefore, each scenario will be just one particular choice of path and repetitions.) Use case testing is a form of black-box system testing. As well as the normal sequences of events, we can test the abnormal sequences that have been speciﬁed in the use case, i.e. how the system should behave when things don’t go according to plan. For example, an attempt to rent a car will usually be successful but we may have decided to specify what should happen in the rare case that the customer owes us money from a previous transaction. In a multi-user system, it is also important to execute use case tests in parallel, to simulate what will happen after deployment. Use case testing is particularly important in development driven by use cases, as described in this book. Since our development is incremental, based on the prioritizing of use cases, use case tests should be run incrementally.

13.3.6 Component Testing A component, in general terms, is a logically separate piece of software, with a clean, well-deﬁned interface. By analogy with hardware components, a software component is independent of the world around it and keeps its internals hidden. If we want to make use of a component, we’re restricted to using its interface, we can’t ‘break it open’ – this shields us from the internal complexity, while still allowing us to connect components to construct something bigger. Any of the following can be considered as a component: an object (perhaps with other objects inside); any group of collaborating objects; a subsystem; an entire system. On that last point, thinking of a system as a component allows us to link systems together, perhaps at different locations. For example, the iCoot system, which runs inside an individual branch, delivers weekly reports to head ofﬁce via a business-to-business-link – in this case, the systems are entirely separate, treating each other as components.

Types of Test

So, component testing is about running tests against whatever we have decided constitutes a component. As usual, we can choose whether we do black-box testing, white-box testing, or both.

13.3.7 Build Testing A build test is used to ensure that our software can be built successfully. To build software means to collect together all the bits and pieces of the system, compile them, link them together and deploy them, along with any resources that they need (databases, ﬁles that are not actually source code). A build test is sometimes called a smoke test. While developing a new engine, a team of mechanical engineers will start up and rev the engine at frequent intervals; if the engine emits smoke, something is wrong (see Figure 13.3), so the engineers shut the engine down, diagnose the problem and ﬁx it, before doing another smoke test to check the repair. For anything less than a trivial system, we can use build tools to ensure that the construction of the system is automated, consistent and complete. For example, Ant, from the open source Apache project (ant.apache.org), takes as input XML ﬁles written by developers; these ﬁles describe how to construct all the pieces of the complete system. The XML ﬁles also contain information about the dependencies between individual pieces, so that the Ant tool can build them in the correct order: for example, we might specify that a particular set of source ﬁles must be compiled before they can be linked together into an executable. Once the system has been built without failures (such as those caused by syntactical or semantic errors, reported by the compiler), the second stage of the build test is to start the system up and go through some critical scenarios – if these scenarios can be completed without any further failures (such as resources not being found or exceptions being thrown by the program), the test has succeeded. After a successful build test, we know that the code and resources we have implemented up to this point work together, so we won’t have problems when we deploy the ﬁnished system. If the scenarios fail, we know that there are faults that must be dealt with before we continue with the development, since there is no point producing a ﬁnished system that won’t build, deploy or perform its essential functions. It is a good idea to perform a build test frequently, hence the term nightly build. This may seem to be over-enthusiastic, but it is beneﬁcial for a large system with many pieces, especially since the pieces are usually developed by separate people, who may be at separate locations, unable to communicate on a daily basis. Regular build testing is also important because it avoids deferring problems until later, when they will generally have become larger (because of the tendency for the work of individuals to diverge over time). Even for a small system, perhaps with only one developer, regular builds still have the general advantages described above. It is natural for developers to build their own code several times

415

416

Chapter 13

Figure 13.3: An unsuccessful smoke test

a day anyway – a successful build and a few scenario walk-throughs demonstrate that they’re making progress. For medium- to large-scale projects, there will usually be a dedicated build team, responsible for writing the build ﬁles, collecting pieces from developers and performing the build tests. This frees ordinary developers from a lot of administrative activity – indeed, as long as the pieces provided by a particular developer cause no problems, they can forget about the build test altogether. As soon as a build test fails, the build team must identify which developer is responsible for the failure and make sure that they help to ﬁx the associated

Types of Test

faults. (Of course, failures may also be caused by members of the build team, using invalid build ﬁles.) As an incentive to ﬁx problems quickly, some kind of forfeit may be imposed by the development team. These may be humorous, low-impact forfeits, such as having to wear a hat marked ‘Build Buster’ until the problem is ﬁxed, or they may be rather more serious, such as being on call twenty four hours a day, seven days a week to come in and ﬁx any faults that you have caused.

13.3.8 Load Testing While we’re developing our system, however much we test as we go along, we’re unlikely to be able to test its behavior under normal load. To use an analogy with engineering: when building a bridge, we may be able to test the strength of individual girders before we add them to the bridge, but we won’t be able to test the whole bridge under a typical load of, say, a thousand lorries, two thousand cars and one thunderstorm. With software load testing, we take the ﬁnished system and apply the normal, expected load to it – for example, ‘100,000 users trying to examine the contents of their bank account simultaneously’. As well as normal (average) load, we may test the system under high load and maximum load, to check that it can cope under these extreme conditions. Software load testing is, in many ways, easier than the load testing carried out by engineers: if a bridge were to fail, we would have a serious problem, so the engineer has to use techniques such as wind-tunnel tests and computer modeling to make sure that, even under extreme conditions, a collapse is impossible. Even then, the engineer allows the bridge to be opened knowing that their conﬁdence is theoretical – how good can the application of scientiﬁc principles and modeling be, when we didn’t actually test the real thing? With software, on the other hand, we can load-test the real code in deployment conditions, safe in the knowledge that, if anything goes wrong, we can ﬁx the code before it goes live. Another advantage of software load testing is that, if a failure does happen when the system is live, we can re-start the system (some would suggest that this is why software is less reliable than engineered products). In the case of a safety-critical system, such as a ﬂy-by-wire airplane, this situation doesn’t apply, so we may use different principles: if we can demonstrate that our ﬂy-by-wire plane is more reliable than a similar plane under the control of a human being, we have a net gain.

Soak Testing Some faults in software only become evident when the software is used continuously for long periods. For example, a system that doesn’t remove temporary ﬁles as it goes along will eventually run out of disk space, but only if it is allowed to run under load for an extended period of time. Soak testing means running our software continuously under high load, to make sure that it doesn’t exhaust its own resources or exhibit any other cumulative problems.

417

418

Chapter 13

Stress Testing Engineers often test a structure to breaking point, for two reasons: ﬁrstly, to ﬁnd out what the breaking point actually is, regardless of the theoretical limits; secondly, to ﬁnd out exactly what happens when the structure does fail, hopefully discovering that only part of it fails. Software stress testing is the process of overloading a system to the point where it can no longer cope. By doing this, we discover what the limits are and we can see how it behaves beyond the breaking point. When performing stress testing, we may deprive the system of the resources it needs, in order to see what happens: for example, we may deprive it of some of the computing power, memory or disk space that it needs to do its work. Generally, with software, we aim to achieve graceful failure: the system tries to minimize the associated losses and stay running. For example, if we have put together a news Web site that typically handles a million simultaneous requests, an extraordinary news item may result in a hundred million attempts to access the system. In such cases, rather than the whole Web site falling over in a heap, it would be better if 10 million requests were handled as normal and the other 90 million users were informed that the Web site is too busy and they must try again later. With the help of live monitoring, we could be made aware of the problem, so that we could decide whether or not to increase the capacity of the site. Another term for stress testing is negative testing. This is negative in the sense of ‘pessimistic’ because the software is taken beyond its operational envelope into areas for which it was not designed.

13.3.9 Installation Testing Installation testing involves seeing how our software behaves in its operational environment. Since the majority of testing may be carried out in a specially built test environment (the test bed), we have not fully tested the software until we have checked it in the live environment too. This is especially true of software that is intended to be portable, designed to run in any number of environments (multiple operating systems or devices). Other terms for installation testing include platform testing and environment testing.

13.3.10 Acceptance Testing Ultimately, the success of a piece of software depends on whether it satisﬁes the needs of its users. Even if the software has been commissioned by the management for the beneﬁt of the organization, if it is not acceptable to the end users, they won’t use it properly, defeating the original purpose. In the case of software manufactured for sale to third parties, the software will only sell, and continue to sell, if it is effective to those who use it. Naturally, we would prefer our software to be good rather than just adequate. Acceptance testing allows end users to try the software before release, to see if they accept it. During acceptance testing, we may ask the users to perform particular tasks or we may

Types of Test

ask them to try to do what they would normally have done by other means. While the tests are being carried out, we can monitor the users’ effectiveness and then gather feedback. Monitoring may include videoing the users to gauge their visible and audible reactions, and recording their keystrokes and mouse movements (the general principle being ‘less is more’ – the less interaction required to get the job done, the better). Although monitoring and gathering feedback is user-focussed, parts of acceptance testing may take place at a higher level: discussing productivity with managers or asking a psychologist to interpret video recordings, for example. Since full acceptance testing is carried out late in the day, during the testing phase, it can be expensive to modify the software to take account of negative results. Thus, we would normally defer the less critical modiﬁcations to a later increment. The principles of spiral and iterative development, with continuous customer involvement, help us to avoid any serious problems.

13.3.11 Regression Tests Whenever software is modiﬁed, by adding, removing or editing source code, we need to be sure that the new version of the software is at least as good as the previous version. Making changes to code can introduce faults. It is not unusual for ‘ﬁxes’ that are not intended to introduce new behavior to contain errors of their own. Regression testing is the process of ensuring that modiﬁcations have not caused the code to regress, i.e. that the code has not ended up worse than before. Thus, we will usually have a suite of tests that we run against our software after every change – the new version of the software should pass these tests just as well as the old. (Ideally, as well as being just as correct, the new version should be just as effective, fast, secure, elegant, and so on.)

13.3.12 Documentation Tests Documentation is an umbrella term that refers to manuals and training materials. Such documentation is crucial for the success of a system: if no-one can use the software or maintain it in the ﬁeld, there was no point producing it in the ﬁrst place. Documentation should be aimed speciﬁcally at both system administrators and end-users, even if that means that two sets have to be produced. Documentation testing is the process of checking that documentation is correct and effective, with input from other authors, system administrators, end users and instructors.

13.3.13 Testing for Security Security testing is the process of ensuring that the system is secure, but what does secure mean? As outlined in Chapter 8, security has the following aspects:

419

420

Chapter 13

• • • • •

privacy authentication irrefutability integrity safety

Our system must be protected against hackers, malicious third parties intent on mischief or criminal activity. Thus, we must test that it is impossible for invalid or unauthorized information, or software, to get in or out of the system. Ethical hackers are specialists in this area: with our permission, they will attempt to hack into our system on its test bed. Any weak points that the ethical hackers identify can then be ﬁxed before release. Most security testing will be performed as part of system testing, usually during the testing phase. This is because, although each piece of the system should be designed with security in mind, we should concentrate our testing efforts on making sure that hackers (and gremlins) can’t get beyond the system boundary.

13.3.14 Metrics A metric is a measurement that we take to assess the quality or effectiveness of our software. Typical metrics include: • Code coverage: It is a widely held principle that code should only exist in a system if it is used, otherwise we have wasted time developing it and increased our maintenance costs and system size unnecessarily. In the context of code reuse, there is something of a conﬂict here: when we implement a class, we may be tempted to add extra public methods that we suspect will be useful in the future when the code is reused in other systems; however, any extra methods that we add aren’t used by the current system. As a compromise, we would expect to tolerate low coverage within libraries and frameworks, while expecting high coverage for any code that is speciﬁc to the current system. • KLOC: An old-fashioned metric, short for thousands (K) of lines of code. Often, KLOCs were used to measure the productivity of developers: the more lines of code a programmer produced, the more productive they were thought to be. However, a programmer who produces a lot of code quickly may be producing more than is necessary, while a methodical, accurate programmer may produce less code that is of better quality. The best object-oriented developers will try to avoid writing their own code, reusing code from libraries, frameworks and other applications, so the KLOC has become largely redundant. KLOCs are still sometimes used to indicate the relative size of an application or system: as in ‘the XYZ system has 15 000 KLOCs’. • Transaction time: A transaction, as the term is used here, is the handling of an incoming request – examples include business transactions such as ‘buy a book’ and database

Automating Tests

•

•

• •

•

transactions such as ‘add a record to the CUSTOMER table’. We can measure the number of transactions handled per second or the average transaction time, as a way of gauging system performance. Depth of inheritance hierarchies: Some would say that an inheritance hierarchy with many levels is too deep. Others would say that there are no rules, it just depends on what you’re doing. It’s fair to say that most inheritance hierarchies, even complex ones, only stretch to half a dozen levels. Therefore, if we spot an inheritance hierarchy with 50 levels, it merits further investigation. Breadth of inheritance hierarchies: This measurement relates to the number of classes across a hierarchy. As with depth, this is a gray area. However, we might conclude that a hierarchy with only three levels but 50 classes at each level is in need of reorganization. Size of methods: A method that is more than, say, 50 lines, should probably be split into several smaller methods: each one would be easier to write and easier to maintain. Degree of coupling: Components with many links between them, passing many messages back and forth, are said to be tightly coupled. This is largely a maintenance issue, because changes to the interface of one component will have a large impact on the other components and vice versa. We should use a greater number of loosely coupled objects or, perhaps, write the system in a client–server style. Degree of cohesion: A cohesive object has a single, coherent, set of responsibilities. For example, an object that represents a pizza on a menu and an actual pizza that someone has ordered has weak cohesion: parts of the system would use the pizza exclusively as a description from the menu, while others would use it in relation to one particular customer. Another example of weak cohesion would be a collection that could be used as a set (unordered, without duplicates) or as a list (ordered, with duplicates). Strong cohesion encourages simplicity, ease of maintenance and reuse.

Most metrics are subjective, so they must be interpreted carefully by humans. Ultimately, the most important metric is ‘Does the software achieve what was set out in the requirements?’

13.4 AUTOMATING TESTS You may have realized by now that there is a lot of work involved in testing software thoroughly. You should also have realized that testing is essential, in order to ensure code quality, correctness and security. However, if it is too time-consuming, developers may be reluctant to do it, especially those who are not actually members of the testing team. Therefore, the more we can automate the process of testing, the better. The ideal is for the tests to be implemented separately from the system software and to run at the touch of a button. Test automation software includes:

421

422

Chapter 13

• Loading tools that simulate the load that the system can expect to be subjected to when it goes live. Tools can allow us to load servers in a client–server system by simulating multiple simultaneous client access; user interfaces by simulating the actions of a person; processes by simulating external access; and peers in a distributed system by simulating access by other peers. • Testing frameworks that comprise reusable code and techniques. The frameworks take the drudgery out of running tests and allows the test developer to concentrate on writing the tests themselves. • Performance monitoring tools that can monitor anything, from the virtual memory in use by an individual process to the network trafﬁc between machines. • Metrics-gathering tools that gather metrics from source code or running programs. For example, we could use such a tool to analyze source code, looking for common faults and examining code quality (in much the same way as a style checker in a word processor). • Speciﬁcation-exercising tools that take an informal speciﬁcation of the interface of our software (written by test developers in some special-purpose, executable language) and attempt to test whether the underlying software conforms to the speciﬁcation. For example, if we specify that a particular message always returns a value less than 10, the tool could verify that this is true for a number of randomly-generated parameters. • Assertion checking tools that check the assertion code inserted by the developer. Assertion checking involves executing expressions at run-time to make sure that they do not fail (evaluate to false). Thus, assertions can be used to test the code every time it is run. These tools (and others) are distinguished by the fact that, once they have been conﬁgured, the tester can start them up and then just sit back and wait for the results.

13.5 PREPARING FOR TESTING From the point of view of a project manager, the major components of a testing strategy are: • A test plan: A statement of how we’re going to achieve the objectives of a system that is correct and meets the customers’ needs. A draft test plan should be written before the development effort gets underway, addressing all phases of the development and all the artifacts produced. A test plan will cover large-scale issues such as testing philosophy and testing techniques, all the way down to tables of individual tests that specify the test name, description, procedure and expected results. • A test bed: An environment that is similar to the live environment, constructed for the sake of testing. For example, if we’re implementing a client–server system to work with clusters of Unix servers (containing the business logic), a database server running on

Preparing for Testing

•

•

•

•

•

a mainframe, and multiple clients on arbitrary platforms, we could construct a test bed consisting of: – ten client machines, using a variety of operating systems; – three Unix servers; – a mainframe; – a TCP/IP-based LAN. A test bed allows us to test the software safely before it is deployed in the live environment, which is much more sensitive. A test harness: A tool that helps us to run our tests or to develop them. A harness supports the testing process, so that our tests are not disorganized. (We could also say that a harness helps us to attach our tests to the code being tested.) For example, a harness may comprise a software framework that provides the code for running our tests and a tool for launching them: the tester only has to implement the tests. Test cases: Logical units each comprising a number of tests, designed to check one particular aspect of our software. For example, we might have a test case based on a logical piece of software – class, package, subsystem – or on a logical piece of the business – buying a book, viewing a bank statement, issuing a query. Test suites: Collections of related test cases. We can collect test suites into higher-level test suites. For example, we might collect the ‘Buy a Share’ and ‘Sell a Share’ test cases into the ‘Trading’ test suite; we might then combine the ‘Trading’, ‘Banking’ and ‘Assets’ test suites into the ‘Money Management’ test suite. Test procedures: Instructions (for people) on how to carry out each test. This will typically involve steps such as: 1. Set up the test. 2. Perform the test (comparing expected and actual results). 3. Report the test result. Each of these will comprise smaller steps, such as ‘Select File→New from the menu bar’ and ‘Enter a ﬁle name’. Test data: Data that we synthesize for the sake of running tests. Before our system is deployed, we do not have real world data available, so test data is our best guess as to the kind of information our system will eventually have to handle. Test data includes data in ﬁles; data in databases; and data passed into the system by typical users. Some form of test data is essential, because we can’t test our software without it: for example, we can’t test a component that is designed to retrieve data from a database unless there is something for it to retrieve. Normally, test data has to be hand-crafted by developers, based on intuition and experience, although we may be lucky enough to have it available from past projects, ideally collected from previous versions of the system in use by the real users.

423

424

Chapter 13

13.6 TESTING STRATEGIES You may think that software should be implemented and then tested, but this strategy is inadequate, for the following reasons: • It takes no account of the other development artifacts, such as use cases and UML diagrams. • It encourages developers to defer fault-ﬁnding to the testing phase, when faults will be more expensive to ﬁx. • It’s unnatural: every programmer wants to try out pieces of code that they have written as soon as they can (it gives them a warm feeling of security and progress). Therefore, what we need is a testing strategy that is continuous: everything we produce is tested, all of the time. Since we’re dealing with different kinds of artifact, our strategy must also involve different kinds of testing (testing by hand, testing with sample code, automated testing, and so on). We do still need an explicit testing phase, because developers, no matter how brave and enthusiastic they are, won’t try to break their own artifacts (their goal is always to ensure that ‘it looks okay so far’). For non-code artifacts, testing involves manual checking by: • The development team, since they’re experts with an interest in success. • Customers, because customers are domain experts who must be involved if we’re to deliver a correct and useful system. • Peers, colleagues who are developers but who are not directly involved in the current project (and who therefore will be less possessive about the artifacts). For code artifacts, testing involves: • The development team, so they can witness their own progress (and produce a better quality product). • Peers, who take part in code reviews, helping to spot faults and opportunities for refactoring. • The testing team, who seek to verify the correctness of the code and validate its effectiveness, for the good of all.

13.6.1 Testing During Development During development, we should test all non-code artifacts manually. If we’re using a tool that is tuned to a particular methodology, it may have traceability features that help us to check that dependent artifacts are consistent. However, there will still be a manual

Testing Strategies

element: there is only so much a tool can do and we may have good reasons for disabling or relaxing automatic traceability (such as personal preference, experience, project guidelines or in-house adaptation of a particular methodology). For the code itself, it is a good idea to encourage programmers to implement tests in parallel with their coding activity. A testing framework such as JUnit can help with this: before writing a piece of system code, the programmer writes some sample code that will test the new piece. Thus, when the new piece is complete, the programmer can verify immediately that they were successful. JUnit programmers perform unit testing, regression testing and, to a lesser extent, integration testing. Programmers should also be encouraged to add dynamic assertion checks to their code. This will enable continuous testing of class invariants, method preconditions and method postconditions, as deﬁned in the speciﬁcation phase. Programmers may also add assertions within methods to check for other common faults, such as branches that should be impossible to reach. Peer reviews should be organized on a regular basis, at the end of each development spiral, for example. All the artifacts under review are distributed to members of the development team and other peers and then a meeting is arranged to go through them in detail, looking for faults and inadequacies. The goal is to certify or sign off the artifacts before proceeding. How formal the peer review process is depends on: company policy; customer policy (for example, defense projects tend to be more stringent); project policy; the stage of the development we have reached (are we still feeling our way or is this a mature project?). With customer reviews, we ask the customer to help us validate and verify the artifacts. Customers should not be involved in reviewing complex artifacts such as sequence diagrams or code. Peer reviews and customer reviews should not be combined, otherwise peers will be inhibited in their constructive criticism by the presence of customers and customers will have to sit through detailed technical discussions. After the ﬁrst increment, all increments should be subject to regular (nightly or weekly) build testing.

13.6.2 Testing During the Testing Phase During the testing phase, the testing team takes over the veriﬁcation and validation of the software before release. This will normally involve: • Re-running the developers’ tests. • Subsystem testing (including layers), based on the design and speciﬁcation phases. • System testing, based mainly on the system use cases developed during the requirements phase. • Acceptance testing, with end users and system administrators. • Installation testing, on a variety of platforms if appropriate.

425

426

Chapter 13

• Documentation testing, of manuals and training materials. • Beta testing. • Metrics gathering, both performance metrics and coding-style metrics. Members of the test team should perform black-box testing and white-box testing.

13.6.3 Testing After Release After release, the testing continues. Every time an end user or administrator starts up or interacts with part of the system, they test it again. There should be a reporting system in place, so that any failures that occur during live use of the system can be reported to the project team (via a support hot-line, e-mail, or a Web site, for example). The reporting system can also be used to provide feedback, such as suggestions for improvement or requests for new features. Between increments, the project team should ﬁx faults as they’re identiﬁed. It is then a project management issue to decide whether ﬁxes are released before the next increment. Normally, in order to reduce workload, a number of ﬁxes will be combined into a single ﬁx pack that customers can use to update their installation. Improvements, and new features, should be deferred until the next increment. Any ﬁx that is applied to the system between increments should be regression-tested.

13.7 WHAT TO TEST FOR Since the real world is unpredictable, we can’t hope to test everything that might happen to our software. Also, since the number of possible inputs and results grows exponentially with the size of the system, we don’t have time to test all the inputs and outputs that we do know about. Therefore, our test cases must be based on experience and guess-work: we must use our experience of the kinds of things that typically go wrong and we must predict which tests will cover the most possibilities – for example, a test that follows ten paths through the system is more useful than a test that follows one path. Another way of thinking about test design is to consider that a test that ﬁnds no faults is a bad test: we would prefer our tests not to skirt around likely problems, because that would leave us with a false sense of security. The kind of testing we do will vary according to the type of application we’re dealing with (desktop application or server, for example) and the application domain (e-commerce system, embedded system, safety-critical system, and so on). Therefore, it is not possible to write down the kinds of tests that should be performed for every system. However, below are a few ideas to get you started (some of these ideas relate to black-box testing, some to white-box testing and some to both):

What to Test For

• Test that the code conforms to its speciﬁcation: We have two main forms of speciﬁcation: use cases and class-based speciﬁcations. These state what the result of correct inputs should be, so we can use them as a guide for designing test cases. We should also test any abnormal paths included in the speciﬁcation (for example, the speciﬁcation may state what should happen if an incorrect password is entered). It is not generally feasible to test for all situations where preconditions are violated, because the test space would be inﬁnite. Instead, we should use application ﬁrewalls to make sure that preconditions are always respected. • Test boundary conditions and mid-range conditions: Often, code is expected to deal with ranges. If, for example, a product id is four to ten characters long, we should test, at a minimum, that the code works for a product id of four characters, for ten characters and for somewhere in-between, say seven characters. • Test for off-by-one faults: With an off-by-one fault, the developer has gone too far or not far enough. For example, when using a for loop to iterate ten times, in Java we might write for (i=1; i<10; i++) when what we actually meant was for (i=1; i<=10; i++). The ﬁrst version of this loop iterates only nine times. • Test special cases: Sometimes, developers are so concerned about what happens 99% of the time, they forget to deal with the other 1%. Special cases often need special test coding but exceptional cases are just as important as normal cases. • Test for unusual situations: For example, when designing an HTML-based client, the developer may assume that the client will only issue one request at a time. But, what if they click a Submit button and then, because they don’t receive the expected result, they click the Back button on the result page? The Web application on the server receives two identical requests from the same browser session, when only one was intended. • Test for memory errors: When using a language such as C++, memory faults are common, because the language allows direct access to memory and because it is more difﬁcult to program. There is also the classic memory leak fault, where memory is consumed by the application but never reclaimed. Even languages like Java that have an automatic garbage collector can suffer from high-level memory leaks, such as adding an object to a collection and forgetting to remove it when it’s no longer needed. • Test for code-copying errors: When a multi-branch statement (such as an if–then–else) has many similar branches, a programmer may copy a branch, paste it multiple times and then hand-edit each one; it’s all too easy to forget to edit one of the branches. • Check the use of operators and precedence: Programmers can confuse similar-looking operators, such as & and &&, or make incorrect assumptions about precedence, for example, + being evaluated before a subsequent *. • Check correct use of equality and identity: As described in Chapter 2, equality and identity have subtle but important differences. For example, are two strings the same string or are they two strings with the same characters?

427

428

Chapter 13

• Test that parameters are passed in the correct order: For example, with a method such as bar(int,int), it would be easy for the programmer to write bar(34,2) when they actually meant bar(2,34). • Check for unintended inﬁnite loops and inﬁnite recursion: An inﬁnite loop is one that never terminates; an inﬁnite recursion is a method that calls itself with no exit condition. • Test for unreachable code: A programmer can inadvertently implement a branch, in an if statement for example, that is impossible to get to: in this case, they have probably made a mistake in the conditional part of one of the if or else parts. • Check that errors are reported: Programs occasionally encounter situations that can’t be handled; in such cases, we need to make sure that an error message is printed to an error log, for further investigation. • Test for calculation faults: It is common to make incorrect calculations based on separators. The analogy used is that of constructing a fence: if we’re building a 16-meter fence, using 4-meter fence panels separated by fence posts, how many fence posts do we need? The correct answer is not (16/4) but (16/4 + 1), because we need an extra post at the end, as illustrated in Figure 13.4. 4m

post

fence

Figure 13.4: Getting the right number of fence posts

• Test application ﬁrewalls for robustness: Having fault-checking code within our system can be wasteful, especially in cases where there can be no fault. For example, if we have a method foo(int) that requires a parameter between 1 and 10, it is wasteful to check the parameter inside the method if we call the method using foo(2). A better approach is to add an application ﬁrewall – a layer of code around the application that ensures that invalid requests can’t get in the ﬁrst place. It’s vital to test that our application ﬁrewalls protect the application, by trying to break through them. • Test correct initialization: When a piece of software is initialized (for example, a system is started or an object is created), it should initialize itself to a sensible state. For example, when we start a server, it should be ready to receive requests from clients immediately after initialization: we should not have to start it and then issue several commands before it can be used. • Test that variables are initialized: When we introduce a variable into our program, we should make sure that it is given a sensible value before it is used. Some programming

What to Test For

languages automatically initialize variables to a sensible default, such as 0 (for numbers) or null (for pointers). However: some languages do not do this and even if they do, the default value may not be what the programmer wanted. • Test that an object complies with its state machine: A piece of software often has a speciﬁc set of legal states. For example, an object that represents a ﬁle may have the following legal states (recorded in the attributes): ‘doesn’t exist in the ﬁle system’; ‘exists and is open for reading’; ‘exists and is open for writing’; ‘exists but is closed’. We would want to check that our ﬁle object can never enter an illegal state such as ‘doesn’t exist and is open for reading’. State models, as recorded in UML state machine diagrams, can help here. • Test that code uses data structure theory: Much has been written about how to implement and use data structures (such as lists, trees, compositions, sets, collections, arrays). For example, if we wish to search for a value in an alphabetical list, it’s wasteful to start at the beginning and go through the elements one after the other. Instead, we should look in the middle ﬁrst: if the value we’re seeking is bigger, we look half way towards the end; if the value we’re looking for is smaller, we look half way towards the beginning, and so on. For a list with 32768 values, a straight linear search takes 16384 (that is, 32768/2) comparisons, on average, to discover whether the value is present; the binary chop search just described only requires 15 comparisons, on average. Familiarize yourself with data structure theory and test that your programmers have too. (If you’re really keen, investigate complexity theory.) • Test for previous faults: After release, every fault that is detected identiﬁes a new test case (because you want to be sure the fault doesn’t return after you’ve ﬁxed it and because it may be an example of a common error). Try keeping a record of all the faults that you and your colleagues have ever discovered, as a guide to the kind of problems that are worth testing for. What about the tests themselves? Should we test them too? If we did, we would have to test the tests of the tests, and the tests of the tests of the tests, and so on. Luckily, the code that is being tested tests the tests: whenever a test fails, we know that there is a fault in the code being tested or a fault in the test itself; during our investigations, we discover which has the fault (and ﬁx it). Also, the tests themselves should be subject to peer review. Case Study A test plan for iCoot A complete test plan for a system the size of iCoot would be a large document, certainly too large to reproduce here. Appendix B.7 contains an outline test plan, of the kind that might be produced early in the ﬁrst increment.

429

430

Chapter 13

13.8 TEST-DRIVEN DEVELOPMENT In the mid-1990s, Kent Beck and Erich Gamma designed a framework for test-driven development in Smalltalk. Dubbed SUnit, the framework proved to be so popular that it now comes in many ﬂavors: Java, C++, C#, VisualBasic, Eiffel, etc. The term xUnit is used to refer to all of these ﬂavors, even though none of them is actually called xUnit (the Java variety is called JUnit, there is a C++ variety called CPPUnit, and so on). As the name suggests, the framework is designed for unit testing, where each unit is a class, or a combination of classes. The xUnit family is a simple family: the designers knew that if it were complicated it would hardly ever be used. The idea is that we write testing code in parallel to the system code; the test code can be run easily (and therefore frequently) by developers. Within the test code, the programmer can write methods that test aspects of the application code – sample programs, if you like. Each method can create objects, check that they have been initialized properly, connect them together, send them messages and check the results, and so on. The core of the framework is illustrated in Figure 13.5. This is an example of the Composite pattern, as described in Chapter 11. The composition allows us to build a hierarchy of tests, with TestSuite objects at the nodes and TestCase objects at the leaves.

<> Test

*

+run(:TestResult)

TestCase

TestSuite

+run(:TestResult)

+run(:TestResult) +addTest(:Test)

MyTestCase +testFoo() +testBar()

MyTestSuite +suite():Test

Figure 13.5: A class diagram for xUnit

*

An Example of Test-Driven Development using JUnit

Each test case comprises a number of test methods with the actual testing code inside. Each test suite is a logical grouping of tests, whatever we think makes a sensible hierarchy. For example, we might gather all the test cases for our business objects into one test suite and combine that into another test suite with a suite containing all of our server layer test cases, and so on, all the way up to the system itself. Once we have implemented our composition, we can use the framework’s tools to run the tests.

13.9 AN EXAMPLE OF TEST-DRIVEN DEVELOPMENT USING JUNIT The best way to understand how test-driven development works is to go through an example. The following narrative starts part way through the development of the iCoot business code. For simplicity, this discussion only relates to the business code, rather than any layers above or below. You may think that this would make the narrative artiﬁcial. However, in keeping with best practice, the business code should be independent of higher layers anyway, so the steps described here are realistic. In addition, the persistence code would eventually be encapsulated by the business objects so, since we will only be testing the interfaces of the business objects, we will suffer no loss of realism in this respect either. (In the ﬁnal system, the only major difference in the interface turns out to be that clients use factory methods to create business objects, rather than the actual constructors.) Java is used as the language for code fragments. However, there is no complicated Java used here (anything unusual, with respect to other object-oriented languages, is only used where necessary and is clearly identiﬁed and explained). Since the code fragments are written in Java, the testing framework is, naturally, JUnit. Let’s start by looking at the objects that we’re going to develop in the design-level class diagram shown in Figure 13.6: • Store holds the address of this location and the cars that we have available for rent. (The Store class comes from the full Coot system rather than iCoot – it is introduced here because it helps with the discussion.) • Address is a simple address class for recording the location of this particular branch. • Car represents a particular instance of CarModel that we own. (For simplicity, we have omitted the CarModelDetails.) • CarModel is the intangible complement to Car: it represents a particular model of car that we have available for rent. • Category indicates the type of car (such as, sports, family, luxury).

431

432

Chapter 13

-address 1

1

Store

Address

-Store(:Address) <> +getAddress():Address +addCar(:Car)

-Address(:String,:String,:String,:String,) <> +getHouse():String +getStreet():String +getCounty():String +getPostcode():String

1

*

-cars:Set CarModel Car -CarModel(:Category,:Make,:String,:int) <> +getCategorey(): Categorey +getMake():Make +getModelNumber():String +getPrice():int

-Car(:CarModel,:int) <> +getModel():CarModel +getTraveled():int +setTraveled(:int) *

1

1

*

-model 1 Make

-makes:List 1..*

-Make(:String) <> getName():String

-category

Category -Category(:String) <> +getName():String

Figure 13.6: Part of the design class diagram for Coot

• Set is a Java collection class that allows us to hold on to any number of unique objects. • Make represents a car manufacturer, identiﬁed by name. Development is already well under way, with Address, CarModel, Category and Make already developed and tested. For each class, a test case has been written to conﬁrm that it functions correctly: AddressTestCase, CarModelTestCase, CategoryTestCase and MakeTestCase, respectively. It is good practice to have one test case per class, designed speciﬁcally to exercise that class (because a class is a unit in its own right). Later, we can add test cases that exercise the classes in combination. The classes have been placed in a package called com.nowherecars.business. The test cases have been placed in the same package as the business classes, so that the tests can use the constructors, which have package visibility.

An Example of Test-Driven Development using JUnit

The tests themselves must be public if we wish to use them with a JUnit test-runner, because the JUnit classes are not in the com.nowherecars.business package. The fact that the tests are visible to clients of the package is not necessarily a bad thing: after all, in the spirit of teamwork and openness, we would be happy for other members of the team to run our tests (especially since we won’t release the code until all the tests have been passed). If we want the tests to be invisible to other clients when the code is deployed, we can simply omit the corresponding classes from the deployed version.

13.9.1 Testing the Car Class Before reaching for our coding tools to start writing Car, we remember that the ultimate skill in test-driven development is to write the tests before the code. Thus, we decide to implement CarTestCase ﬁrst. Writing a test case with JUnit is simple: the new class must inherit from the JUnit TestCase class and it must contain one or more methods that begin with test – the test runner automatically looks for such methods and executes them as part of the test case. Remembering that each test method should test one thing and test it thoroughly, we decide that our ﬁrst test will check the creation of a Car. So, our initial CarTestCase is as follows: public class CarTestCase extends TestCase { public void testCreate() { } }

Inside the testCreate method, we will create a Car, using the constructor which takes a CarModel and the distance traveled on the clock. The creation of the CarModel requires creation of a Category (with a name) and a List of Make objects (each with a name). For convenience, we will only add one Make to the CarModel – after all, we’re not testing the CarModel implementation at this point. We will record each attribute as a local variable, both for clarity and in case we need to access the value later. After writing the code for creating the objects, our testCreate method looks like this: public void testCreate() { Category category = new Category("Saloon"); Make make = new Make("Fort"); List<Make> makes = new LinkedList<Make>(); makes.add(make); String modelNumber = "Blur 1.6"; int price = 30; CarModel carModel = new CarModel(category, modelNumber, makes, price); int traveled = 234243; Car car = new Car(carModel, 33445); }

433

434

Chapter 13

This code fragment uses a Java template class (List). The template type (<Make>) tells the Java compiler what kind of object will be in the collection, so that we can’t add anything else. Now that we’ve created a Car, we need to do the real work of the test: checking that it was created correctly. In order to do that, we will use the getters for the new Car to compare its attributes with the ones that we used during construction. Our test case has inherited a number of assertX methods from the TestCase class that we can use to make declarations about the expected state of the program – if any of these fail, the test runner will produce a failure report giving us details of where and when the failure happened. The methods that we’re interested in here are assertSame, which takes two objects as parameters and checks that they’re identical, i.e. that they both refer to the same object; and assertEquals, which takes two object parameters and checks that they’re equal, i.e. that they have similar attributes or that they’re of equal value. For example, • • • •

assertSame(aCar,aCar)

will pass.

assertSame(new Object(),new Object())

will fail.

assertEqual(new Category(”Sports”),new Category(”Sports”)) assertEqual(10,11)

will pass.

will fail.

For our business layer, when we create a Car, the CarModel that we pass in should be identical to the one returned by the getter. That is to say, our program will use the identity of internal objects to represent the identity of objects in the real world, rather than using attribute values. (An alternative approach would be to allow ‘same attributes’ to indicate ‘same object’, but this would lead to more copying of objects.) Thus, we need to use assertSame to check the correct setting of the CarModel. On the other hand, the distance traveled is a primitive value (a number of kilometers), so it should be tested using assertEqual. Therefore, we add two new lines to the end of testCreate: ... assertSame(car.getModel(), carModel); assertEquals(car.getTraveled(), traveled); }

13.9.2 Implementing the Car Class Having implemented the ﬁrst test for the Car class, we turn our attention to the implementation of Car itself. Car needs: ﬁelds for the car model and the distance traveled; a package constructor that takes initial values for both ﬁelds; getters for each ﬁeld; a setter for the distance traveled (we must be able to change the number of kilometers on the clock, so that we can update it when a customer returns the real car, but it wouldn’t make sense to be able to change the model). Our Car class is shown below (package access is the default in Java, so the constructor needs no visibility speciﬁer):

An Example of Test-Driven Development using JUnit

public class Car { private CarModel model; private int traveled; Car(CarModel m, int t) { model = m; traveled = t; } public CarModel getModel() { return model; } public int getTraveled() { return traveled; } public void setTraveled(int t) { traveled = t; } }

Eager to try out our new test as soon as possible, we start up the JUnit test runner, select our test case and hit the Run button. To our surprise, the test runner indicates that there were failures (as shown in Figure 13.7) and the bar is ﬁlled with red. Looking closely at the information given by the test runner, we can see which kind of assertion failed (a numerical comparison between 234243 and 33445), and also where the failure happened. Looking at the code for testCreate, we quickly discover that the fault is in the test code and not in the business code at all: during our coding of the test, we introduced a local variable called traveled, but we didn’t use it when creating the car, so the assertEquals comparison failed. Once we have replaced the number 33445 with traveled, we recompile and re-run the test (which, conveniently, doesn’t require a re-start of the test runner). This time, the bar ﬁlls up with green, the test runner’s way of indicating that everything’s ﬁne (see Figure 13.8).

13.9.3 Refactoring Tests We’ve now completed one test-driven development cycle: write a test, write the code, run the test, ﬁx any problems and run the test again. To complete the testing of the Car class, we need to exercise the setTraveled method. As before, we decide to write the test ﬁrst, adding the following testSetTraveled method to CarTestCase: public void testSetTraveled() { Category category = new Category("Luxury"); Make make = new Make("Plexus"); List<Make> makes = new LinkedList<Make>(); makes.add(make);

435

436

Chapter 13

Figure 13.7: An unsuccessful JUnit test run

String modelNumber = "Neo STS 3.0"; int price = 109; CarModel carModel = new CarModel(category, makes, modelNumber, price); int traveled = 32432; Car car = new Car(carModel, traveled); int newDistance = 534; car.setTraveled(newDistance); assertEquals(car.getTraveled(), newDistance); }

We run this test against the Car class and everything works ﬁne, the test runner conﬁrming that two tests have been run. However, we notice that testSetTraveled has a lot in common with testCreate – not surprising since they must both start by creating a Car. For our purposes, it would be quite reasonable for both test methods to use cars with similar attributes. Therefore, we decide to pull out the similar code and place it in a shared method. This process is called refactoring: reorganizing code to improve quality.

An Example of Test-Driven Development using JUnit

Figure 13.8: A successful JUnit test run

We could call the shared method explicitly at the start of each test, but the JUnit designers have already thought of that: if we add a method called setUp, it is called automatically before each test method is run. (If we deﬁne a tearDown method, it is also called automatically, after each test – we can use this to release any external resources used during the tests.) So, what we need is a setUp method that looks like this: protected void setUp() { category = new Category("Saloon"); make = new Make("Fort"); makes = new LinkedList<Make>(); makes.add(make); modelNumber = "Blur 1.6"; price = 30; carModel = new CarModel(category, makes, modelNumber, price); traveled = 234243; car = new Car(carModel, traveled); }

437

438

Chapter 13

Because of the way JUnit has been designed, the setUp method has protected rather than package visibility, which would allow it to be invoked by subclasses and by other classes in the same package. In practice, we would not expect test developers to do such a thing. Our shared ﬁelds will be created before each test; this may seem wasteful, but it’s actually good practice, because it guarantees that our tests are independent of each other. Before we proceed, we need to add ﬁeld declarations for category, makes, modelNumber, price, carModel, traveled and car, so that they can be accessed inside any test method. Therefore, we add the following declarations to our class: private Category category; private List<Make> makes; private String modelNumber; private private private private

int price; CarModel carModel; int traveled; Car car;

Our testCreate and testSetTraveled methods can now be much simpler: public void testCreate() { assertSame(car.getModel(), carModel); assertEquals(car.getTraveled(), traveled); } public void testSetTraveled() { int newDistance = 534; car.setTraveled(newDistance); assertEquals(car.getTraveled(), newDistance); }

Having made the changes, we re-run our tests and everything still works: as well as checking our business code, the JUnit framework helps us to make sure that modiﬁcations to our tests are correct, a form of regression testing. Having completed our implementation and testing of CarTestCase and Car, the ﬁnished classes look like this: public class Car { private CarModel model; private int traveled; Car(CarModel m, int t) { model = m; traveled = t; } public CarModel getModel() { return model;

An Example of Test-Driven Development using JUnit

} public int getTraveled() { return traveled; } public void setTraveled(int t) { traveled = t; } } public class CarTestCase extends TestCase { private Category category; private List<Make> makes; private String modelNumber; private private private private

int price; CarModel carModel; int traveled; Car car;

protected void setUp() { category = new Category("Saloon"); make = new Make("Fort"); makes = new LinkedList<Make>(); makes.add(make); modelNumber = "Blur 1.6"; price = 30; carModel = new CarModel(category, makes, modelNumber, price); traveled = 234243; car = new Car(carModel, traveled); } public void testCreate() { assertSame(car.getModel(), carModel); assertEquals(car.getTraveled(), traveled); } public void testSetTraveled() { int newDistance = 534; car.setTraveled(newDistance); assertEquals(car.getTraveled(), newDistance); } }

13.9.4 Creating a Test Suite for Regression Testing Our last task, according to the class diagram we saw earlier, is to implement and test the Store class. As usual, we start with the testing code, producing a StoreTestCase class. Here’s how it looks after the implementation of testCreate:

439

440

Chapter 13

public class StoreTestCase extends TestCase { private Address address; private Store store; private CarModel carModel; private Car car; protected void setUp() { address = new Address("9", "Ash Lane", "Greater Manchester", "SK4 3HJ"); store = new Store(address); Category category = new Category("Vintage"); Make make = new Make("Mostin"); List<Make> makes = new LinkedList<Make>(); makes.add(make); String modelNumber = "Wheely 1950"; int price = 89; carModel = new CarModel(category, makes, modelNumber, price); int traveled = 435345; car = new Car(carModel, traveled); } public void testCreate() { assertSame(store.getAddress(), address); } }

This time around, we’ve anticipated the need for a Store in multiple tests, so we’ve written a setUp method immediately. Our initial Store class looks like the following (notice how the task of adding a Car is delegated to the private Set): public class Store { private Address address; private Set cars; Store(Address a) { address = a; } public Address getAddress() { return address; } public void addCar(Car c) { cars.add(c); } }

When we run our ﬁrst test, everything works ﬁne. However, we have now built up a number of test cases – MakeTestCase, CategoryTestCase, CarModelTestCase, CarTestCase – each

An Example of Test-Driven Development using JUnit

of which has to be run individually. This is inconvenient, especially since we should rerun all related tests whenever a piece of code changes. What we need is a test suite – a group of test cases and other test suites that can be run as one. To build a test suite in JUnit, we add a class that inherits from TestSuite and give it a method called suite, which returns all the tests in the suite. Here’s how ours looks: public class CootBusinessTestSuite extends TestSuite { public static Test suite() { TestSuite result = new TestSuite(); result.addTestSuite(MakeTestCase.class); result.addTestSuite(CategoryTestCase.class); result.addTestSuite(CarModelTestCase.class); result.addTestSuite(CarTestCase.class); result.addTestSuite(AddressTestCase.class); result.addTestSuite(StoreTestCase.class); return result; } }

The ﬁrst thing to notice here is that suite returns an object of type Test, rather than TestSuite. However, you may recall from Figure 13.5, that TestSuite and TestCase both inherit from Test. By declaring that it returns a Test, the suite method allows us to return a TestCase instead, which means that we can build a hierarchy of test suites and test cases, while the test runner only has to worry about the single type called Test. Another odd thing about the suite method is the way that we add our test cases to the suite: result.addTestSuite(MakeTestCase.class). This is a Java trick that allows the test runner to ﬁnd all methods in the test case that start with test. In order to run all the tests in a test suite, we use the same test runner as before, only now we point it at the CootBusinessTestSuite class instead of an individual test case such as CarTestCase. When we run our suite, we see the green bar again, indicating that our two new classes are okay, and also that every other part of the suite is okay. This is true regression testing: we have checked that adding new code hasn’t broken anything that we already had.

13.9.5 Testing Across Methods Next, we decide to test the addCar method, by providing a testAddCar method in the StoreTestCase: public void testAddCar() { store.addCar(car); assertTrue(store.containsCar(car)); }

441

442

Chapter 13

The assertTrue assertion succeeds only if its parameter is true. Thus, the test will succeed if adding a Car to the Store results in the Store containing that Car – this may seem obvious, but even obvious tests need to be implemented. While writing the assertTrue statement, we have discovered the need for a containsCar method on the Store class itself. Initially, we need this just to check that the addition was successful, but we decide to make it public, because it will probably be useful to other clients. The containsCar method wasn’t included in the class diagram we saw earlier, because we didn’t know that we needed it. This illustrates that there must be some ﬂexibility between design and coding, which is generally considered to be a good thing. The addition we need to the Store class is: public boolean containsCar(Car c) { return cars.contains(c); }

Our new testAddCar method does more than just test addCar: it also tests containsCar. This is common for JUnit test methods. This time, when we run the tests, we get a test failure with a message containing NullPointerException: a sure sign, to a Java programmer, that we’ve forgotten to create an object. This time the culprit turns out to be the Store constructor: we’ve remembered to initialize the address ﬁeld from the parameter, but we’ve forgotten to create the Set ﬁeld that’s used to manage our Car objects. What we need to do is to add the following line to the constructor: cars = new HashSet();

(Don’t worry about HashSet, it’s just one of Java’s concrete implementations for the Set interface.) Now JUnit has helped us spot the incomplete initialization of an object’s innards, something that we might not otherwise spot until much later. Another win for test-driven development.

13.9.6 Completing the Store Class There is one more method to add to Store before today’s assignment is ﬁnished. This method, containsAlternativeCar will return true if and only if there is another Car in the Store that serves the same purpose as the Car given as a parameter: in car rental terms, this means any car that is the same model, regardless of the number of kilometers traveled. First we add a test to StoreTestCase that will conﬁrm that the new Store method works correctly: public void testContainsAlternativeCar() { int traveled2 = 4435;

An Example of Test-Driven Development using JUnit

store.addCar(car); Car car2 = new Car(carModel, traveled2); assertTrue(store.containsAlternativeCar(car2)); }

Finally, here’s the method that we need to add to Store: public boolean containsAlternativeCar(Car outer) { for (Car inner : cars) { if (inner.getModel() == outer.getModel()) { return true; } } return false; }

The contents of this method may look strange if you’re not familiar with Java. It uses a variety of the Java for loop, designed specially for Collection classes, to examine every Car in cars. Within the loop, the model of the inner Car is compared to the model of the outer Car (that was passed in as a parameter); if the models match, we return true. If we examine every Car in the Store without ﬁnding a match, we return false. Happily, when we run our tests this time, the bar stays green, so we know the code is correct (the JUnit motto is ‘Keep the bar green to keep the code clean’). The ﬁnal versions of Store and StoreTestCase are shown below: public class Store { private Address address; private Set cars; Store(Address a) { address = a; cars = new HashSet(); } public Address getAddress() { return address; } public void addCar(Car c) { cars.add(c); } public boolean containsCar(Car c) { return cars.contains(c); } public boolean containsAlternativeCar(Car outer) { for (Car inner : cars) { if (inner.getModel() == outer.getModel()) {

443

444

Chapter 13

return true; } } return false; } } public class StoreTestCase extends TestCase { private Address address; private Store store; private CarModel carModel; private Car car; protected void setUp() { address = new Address("9", "Appletree Avenue", "SK5 9PT"); store = new Store(address); Category category = new Category("Vintage"); Make make = new Make("Mostin"); List<Make> makes = new LinkedList<Make>(); makes.add(make); String modelNumber = "Wheely 1950"; int price = 89; carModel = new CarModel(category, make, modelNumber, price); int traveled = 435345; car = new Car(carModel, traveled); } public void testCreate() { assertSame(store.getAddress(), address); } public void testAddCar() { store.addCar(car); assertTrue(store.containsCar(car)); } public void testContainsAlternativeCar() { int traveled2 = 4435; store.addCar(car); Car car2 = new Car(carModel, traveled2); assertTrue(store.containsAlternativeCar(car2)); } }

Our narrative ends at this point, having covered a real-world example of test-driven development (complete with faults and iterations, as they happened). The developers moved on to add homes, universal identiﬁers, persistence code and server code. Then, a simple

Further Reading

client was added to test everything that had been developed. Throughout, test cases were created and composed into test suites, layer by layer, and the tests were run frequently.

13.10 SUMMARY In this chapter we have looked at: • The terminology used by testers to summarize complex concepts and tasks. Although there are a couple of dozen terms in common use, they’re used consistently by the testing community, so it’s worth knowing what they are. • How to go about testing a large system. This involved discussions of test planning, continuous testing, testing of all artifacts (not just code), and how the different people should be involved. • A worked example of test-driven development. Test-driven development is valuable, but it should never be considered a substitute for the aggressive attempts to break code by colleagues and customers, during the ofﬁcial testing phase.

FURTHER READING A long-standing text on all aspects of software testing is [Myers et al. 04], considered by many to be essential reading for those who wish to understand the theory of testing and how to apply it in the real world. One of the inventors of test-driven development, Kent Beck, describes the philosophy and practice in [Beck 02]. The home page for JUnit is www.junit.org, containing the framework itself, plus technical documentation, articles and examples.

445

Appendix A Ripple Summary This appendix gives a complete overview of Ripple, the simpliﬁed methodology used in this book. Table A.1 summarizes the artifacts that you should produce, by phase. To produce these artifacts, proceed as follows (but remember to spiral, iterate and deliver incrementally): 1. Project Genesis (with Customer) (a) Get an idea of what the customer is looking for or tell them what they need. (b) Get requirements documentation from the customer, as a mission statement or a longer document; if neither of these exist, produce an informal requirements document with the customer. 2. Assigning Responsibilities (a) Decide on the development roles (e.g. planning, management, timekeeping, development, testing, system administration). (b) Decide who is going to be responsible for each role. 3. Producing a Workbook: Produce a paper or on-line workbook, to contain all project artifacts. 4. Producing a Glossary (a) Produce a glossary for recording the deﬁnitions of project terminology. (b) Update the glossary throughout development. 5. Producing a Project Plan (a) Do some initial planning for spirals, phases and increments, and produce a schedule. (b) Review and adjust the project plan at regular intervals throughout development. 6. Producing a Test Plan: Produce a test plan that addresses continuous testing, the testing phase, customer reviews, deployment and maintenance. 7. Business Requirements (with Customer) (a) Produce a business actor list (with descriptions). (b) Produce a business use case list (with descriptions). (c) (Optional) Brainstorm or illustrate business use cases with activity diagrams.

Ripple Summary

Table A.1: Artifacts by phase Phase

Artifacts

UML

447

448

Appendix A

(d) (Optional) Brainstorm or illustrate business use cases with communication diagrams. (e) Produce business use case details. 8. System Requirements (with Customer) (a) (b) (c) (d) (e) (f) (g) (h)

Brainstorm system interaction using user interface sketches. Produce a system actor list (with descriptions). Produce a system use case list (with descriptions). Produce a system use case diagram. Produce a system use case survey. Produce system use case details. Produce supplementary requirements for the system. Produce system use case priorities.

9. Analysis (a) Produce an analysis class diagram. (b) Produce an attribute list (with descriptions). (c) (Optional) Use state machines to model complex entity life cycles, recording results on state machine diagrams. (d) Perform use case realization, documenting results using communication diagrams. (e) Produce an operation list (with descriptions). 10. System Design (a) (b) (c) (d) (e) (f) (g) (h)

Make technology choices. Look for reuse opportunities (libraries, patterns and frameworks). Produce layer diagrams. Write a layer interaction policy. Design the package structure and record it on package diagrams. Produce a deployment diagram. Write a security policy. Write a concurrency policy.

11. Subsystem Design (a) Deﬁne business services. (b) Look for more reuse opportunities (libraries, patterns and frameworks). (c) Map analysis classes to business layer classes: class list (with descriptions), class diagram, ﬁeld list (with descriptions). (d) Produce a database schema. (e) Design classes for other layers (e.g. server and protocol, servlets, control, persistence). (f) Perform business service realization, recording results on sequence diagrams.

Ripple Summary

(g) Produce a message list (with descriptions). (h) Finalize the user interface design. 12. Class Speciﬁcation (a) Produce an informal speciﬁcation for each class. (b) Record the informal speciﬁcation in design and source code. 13. Implementation (a) Write unit tests. (b) Write implementation code. 14. Testing (a) Get the test team to test the system. (b) Fix faults. 15. Deployment (a) Produce manuals and courseware. (b) Install code artifacts on customer’s system. (c) Train customer. 16. Maintenance (a) Fix faults. (b) Incorporate customer feedback, ideas for improvement and market changes into new increments.

449

Appendix B iCoot Case Study

B.1 BUSINESS REQUIREMENTS This section documents the business requirements modeling carried out during the requirements phase of the iCoot development, in terms of project mission statement and business use case model. The business use case model also applies to the full Coot system.

B.1.1 Customer’s Mission Statement Below is the mission statement delivered by Nowhere Cars at the start of the Coot project: Since we automated the tracking of cars at our stores – using bar codes, counter-top terminals and laser readers – we have seen many beneﬁts: the productivity of our rental assistants has increased 20%, cars rarely go missing and our customer base has grown strongly (according to our market research, this is at least partly due to the improved perception of professionalism and efﬁciency. The management feels that the Internet offers further exciting opportunities for increasing efﬁciency and reducing costs. For example, rather than printing catalogs of available cars, we could make the catalog available to every Internet surfer for browsing on-line. For privileged customers, we could provide extra services, such as reservations, at the click of a button. Our target saving in this area is a reduction of 15% in the cost of running each store. Within two years, using the full power of e-commerce, we aim to offer all of our services via a Web browser, with delivery and pick-up at the customer’s home, thus achieving our ultimate goal of the virtual rental company, with minimal running costs relative to walk-in stores. Working with the customer, this mission statement was expanded into business use cases.

B.1.2 Actor List • Assistant: An employee at one of our stores who helps a Customer to rent a Car and reserve a CarModel. • Customer: A person who pays us money in return for one of our standard services.

Business Requirements

• Member: A Customer whose identity and credit-worthiness have been validated and who, therefore, has access to special services (such as making a Reservation by phone or over the Internet). • NonMember: A Customer whose identity and credit-worthiness have not been checked and who, therefore, must provide a deposit to make a Reservation and surrender a copy of their License to rent a Car. • Auk: The existing system that handles Customer details, Reservations, Rentals and the Catalog of available CarModels. • DebtDepartment: The department that deals with unpaid fees. • LegalDepartment: The department that deals with accidents in which a rented Car has been involved.

B.1.3 Use Case List • B1:Customer Rents Car: Customer rents a Car that they have selected from those available. • B2:Member Reserves CarModel: Member asks to be notiﬁed when a CarModel becomes available. • B3:NonMember Reserves CarModel: NonMember pays a deposit to be notiﬁed when a CarModel becomes available • B4:Customer Cancels Reservation: Customer cancels an unconcluded Reservation, by phone or in person. • B5:Customer Returns Car: Customer returns a Car that they have rented. • B6:Customer Told CarModel Is Available: Customer is contacted by an Assistant when a Car becomes available. • B7:Car Reported Missing: Customer or Assistant discovers that a Car is missing. • B8:Customer Renews Reservation: Customer renews a Reservation that has been outstanding for more than a week. • B9:Customer Accesses Catalog: Customer browses the catalog, in-Store or at home. • B10: Customer Fined for Uncollected Reservation: Customer fails to collect a Car that they have reserved. • B11:Customer Collects Reserved Car: Customer collects a Car that they have reserved. • B12:Customer Becomes a Member: Customer provides CreditCard details and proof of address to become a Member. • B13:Customer Notiﬁed Car Is Overdue: Assistant contacts Customer to warn them that a Car they have rented is more than a week overdue. • B14:Customer Loses Keys: Replacement keys are provided for a Customer who has lost them. • B15:MembershipCard Is Renewed: Assistant contacts Member to renew membership when their CreditCard has expired. • B16:Car Is Unreturnable: A Car is wrecked or breaks down.

451

452

Appendix B

B.1.4 Use Case Communication Diagrams Communication diagrams were not used widely during business requirements modeling (although they were used extensively during system requirements gathering). However, one diagram (see Figure B.1) was produced to illustrate the external and internal actors involved in B3:NonMember Reserves CarModel.

2:find CarModel

1:reserve CarModel Non Member

3:get deposit, License 5:confirm Reservation

Assistant

4:create Reservation

CarModel

Auk Interface

Reservation

Figure B.1: Communication diagram for B3:NonMember Reserves CarModel

B.1.5 Use Case Activity Diagrams Activity diagrams were not used widely during the business requirements modeling. However, one diagram (see Figure B.2) was produced to illustrate the ﬁner points of the B3:NonMember Reserves CarModel use case.

B.1.6 Use Case Details B1:Customer Rents Car.

1. 2. 3. 4. 5.

tells Assistant which CarModel they’d like to rent. If Auk indicates no such Car is available, Customer is offered an alternative. If there is a Car available, Assistant marks the Car as taken in Auk. Assistant asks for Customer’s License to conﬁrm their identity. For a Member, Assistant takes their number from their MembershipCard and checks that they have no outstanding fees and that they have not been barred. 6. For a NonMember, Assistant checks whether they’re already in Auk; if they’re not, Assistant scans a copy of their License into Auk, and records their name, phone number and license number. 7. If Customer’s details are satisfactory and they have paid any outstanding fees, they’re charged for the Rental. 8. If the payment fails, the Car is released in Auk. Customer

Business Requirements

(NonMember) Tell Assistant which CarModel

(Assistant) Ask for Deposit and License

(NonMember) Find deposit

(NonMember) Find License

(Assistant) Find CarModel on Auk

(Assistant) Check deposit and License

[not ok] [ok] (Assistant) Make Reservation

Figure B.2: Activity diagram for B3:NonMember Reserves CarModel

9. If the payment does not fail, the Customer is given the keys and directed to the display area. B2:Member Reserves CarModel.

1. Member tells Assistant their membership number (over the phone or in person). 2. Member tells Assistant which CarModel to reserve. 3. If Member has not been barred, their CreditCard has not expired, and they have no outstanding fees, a Reservation is made on Auk.

453

454

Appendix B

4. If the Reservation is being made over the phone, Member can pay outstanding fees by conﬁrming their CreditCard details, which must match those stored in Auk and must not have expired. 5. Member is told the reservation number. B3:NonMember Reserves CarModel.

1. 2. 3. 4. 5. 6.

tells Assistant which CarModel to reserve. Assistant ﬁnds CarModel on Auk. Assistant asks for a deposit for the Reservation. Assistant asks for NonMember’s License and phone number. Assistant checks License visually. If License looks valid, Assistant creates a new Reservation, recording the License number, phone number and a scan of the License in Auk. 7. Assistant gives NonMember a Reservationslip containing the unique reservation number. NonMember

B4:Customer Cancels Reservation.

1. At any time, Customer can cancel a Reservation. 2. Member objects can do this over the phone or in person, by providing their membership number. 3. NonMembers must cancel in person: they present their License to an Assistant, who checks that it matches the scan in Auk, and refunds their deposit. 4. If a Car has already been moved to the reserved area, a matching Car is moved back to the display area. B5:Customer Returns Car.

1. When a Car is returned to the check-in area, Assistant scans bar code to conﬁrm the return and checks that the tank is full. 2. Car is returned to the display area by an Assistant. 3. If Customer returns an overdue Car or a Car with a tank that is not full, Customer must pay the appropriate amount – Members can do this using their existing credit card details, if they have not expired. 4. If the Customer refuses to pay, their details are passed to the DebtDepartment. B6:Customer Told Car Model is Available.

1. When a Car is returned, Auk tells Assistant whether it matches any Reservation objects. 2. If so, Assistant moves Car to the reserved area. 3. On a ﬁrst-come-ﬁrst-served basis, Assistant will try to contact a matching Customer by phone. 4. If a Customer can’t be reached within two days, their Reservation is canceled and the Car is moved out of the reserved area to the display area.

Business Requirements

B7:Car Reported Missing.

1. If a Car that Auk indicates is in the display area can’t be found when it is needed or during a stock check, Car is reported stolen to the police. 2. If a Car is reported missing by a Customer, it is reported stolen to the police, along with License details of the Customer (as the last known keeper of the vehicle). 3. In both cases, the date of loss is recorded on Auk. B8:Customer Renews Reservation.

1. If a Reservation can’t be satisﬁed within seven days, the Reservation must be renewed. 2. Assistant has two days to contact Customer by phone to see if they wish to renew the Reservation for a further seven days. 3. If the Customer doesn’t wish to renew, the Reservation is canceled; Customer must return to the Store and present their License to retrieve their deposit. B9:Customer Accesses Catalog.

1. Customers can come into the Store to browse a paper catalog. 2. For a fee, they can take a copy of the catalog home with them. 3. If they choose to join the mailing list, they will receive a free copy of the catalog by mail every six months. B10: Customer Fined for Uncollected Reservation.

1. If a CarModel has become available for a particular Reservation and an Assistant told the Customer by phone that it’s available, Customer has two days to collect. 2. If Customer fails to collect, the Reservation is concluded and an Assistant moves a matching Car from the reserved area back to the display area. 3. For NonMembers, their deposit is forfeited. 4. For Members, a ﬁne is recorded on Auk and their details are passed to the DebtDepartment. B11:Customer Collects Reserved Car.

1. 2. 3. 4.

comes to the Store to collect a Car from the reserved area. Customer presents License. If the License matches the details on Auk, the Reservation is marked as concluded. An Assistant gives the keys to the Customer and directs them to the reserved area. Customer

B12:Customer Becomes Member.

1. In order to become a Member, Customer must offer their License, further proof of address, and a credit card. 2. Assistant checks License and proof of address. 3. Assistant checks CreditCard with CreditCardCompany.

455

456

Appendix B

4. If okay, Assistant records License number, address, phone number and CreditCard details in Auk. 5. Auk issues new MembershipCard with unique membership number. 6. If the CreditCard expires, no further member actions are permitted unless the member returns to the Store to show a new CreditCard.

B13:Customer Notiﬁed Car is Overdue.

1. Since a Rental is paid up-front, Customer is warned if they have forgotten to return a Car. 2. If Car is more than one week overdue, an Assistant will attempt to contact Customer by phone. 3. If Customer can’t be contacted for two weeks, Car is reported missing (see B7).

B14:Customer Loses Keys.

1. If Customer notiﬁes Assistant that they have lost keys, replacement keys are provided, by courier if necessary. 2. Cost of replacement is added to Customer’s details in Auk.

B15:MembershipCard is Renewed.

1. 2. 3. 4. 5.

Auk records that Member whose CreditCard has expired is not in good standing. Auk informs Assistant that Member’s Credit card has expired. Assistant contacts Member by phone to tell them that they must renew their membership. Member returns to Store with fresh CreditCard and details are entered into Auk. Auk records that Member is in good standing.

B16:Car is Unreturnable.

1. If Customer tells Assistant that Car is wrecked or breaks down, Assistant arranges recovery. 2. If Car is wrecked, details are passed to LegalDepartment.

B.2 SYSTEM REQUIREMENTS This section documents the results of system modeling during the requirements phase of the iCoot development, in terms of user interface sketches and a system use case model.

B.2.1 User Interface Sketches The user interface sketches for iCoot, produced with the help of the customer, are shown in Figures B.3 through B.10.

System Requirements

5

4

7

6

8

Search Index Membership Rentals Reservations Password

category a category b category c category d manufacturer a manufacturer b manufacturer c manufacturer d

engine size a engine size b engine size c engine size d

Retrieve...

2

Non-Members see only Search + Index pages; Members needlogon and logoff mechanism too.

Figure B.3: User interface sketch 1 (creating a query)

Search Index Membership Rentals Reservations Password

car model a car model b car model c car model d car model e car model f car model g car model h car model i

< Go Back 1

3

Figure B.4: User interface sketch 2 (viewing results)

457

458

Appendix B

Search Index Membership Rentals Reservations Password

Category Sports Make(s) Abc, Def Model 358 Engine Size 3.6 Description Waffle Daily Price €89.50 < Go Back

Reserve...

Advert...

Poster...

2

Figure B.5: User interface sketch 3 (viewing car model details) Search Index Membership Rentals Reservations Password

index entry a index entry b index entry c index entry d index entry e index entry f index entry g index entry h index entry i

Retrieve...

Proceed as for Search page

Figure B.6: User interface sketch 4 (selecting an index heading) Search Index Membership Rentals Reservations Password

Personal Details

Address

Credit Card

< Go Back

Reserve...

Advert...

Poster...

Figure B.7: User interface sketch 5 (viewing membership details)

System Requirements

Search Index Membership Rentals Reservations Password

rental a rental b rental c rental d

Figure B.8: User interface sketch 6 (viewing rentals)

Search Index Membership Rentals Reservations Password

reservation a reservation b reservation c

Cancel...

Figure B.9: User interface sketch 7 (viewing reservations)

Search Index Membership Rentals Reservations Password

Old Password ****** New Password ******** Repeat New Password ******** Change... Clear Fields

Figure B.10: User interface sketch 8 (changing a password)

459

460

Appendix B

B.2.2 Actor List • Customer: A person using a Web browser to access iCoot. • Member: A Customer who has presented their name, address and CreditCard details at one of our Stores; each Member is given an Internet password to accompany their membership number. (Specializes Customer.) • NonMember: A Customer who is not a Member. (Specializes Customer.) • Assistant: An employee at a Store who contacts Members to tell them about the progress of their Reservations.

B.2.3 Use Case List • U1:Browse Index: A Customer browses the index of CarModels. (Specializes U13, includes U2.) • U2:View Results: A Customer is shown the subset of CarModels that were retrieved. (Included by U1 and U4, extended by U3.) • U3:View CarModel Details: A Customer is shown the details of a retrieved CarModel, such as description and advert. (Extends U2, extended by U7.) • U4:Search: A Customer searches for CarModels by specifying Categories, Makes and engine sizes. (Specializes U13, includes U2.) • U5:Log On: A Member logs on to iCoot using their membership number and current password. (Extended by U6, U8, U9, U10 and U12.) • U6:View Member Details: A Member views some of the details stored by iCoot, such as name, address and CreditCard details. (Extends U5.) • U7:Make Reservation: A Member reserves a CarModel when viewing its details. (Extends U3.) • U8:View Rentals: A Member views a summary of the Cars they’re currently renting. (Extends U5.) • U9:Change Password: A Member changes the password that they use to log on. (Extends U5.) • U10: View Reservations: A Member views summaries of their unconcluded Reservations, such as date, time and CarModel. (Extends U5, extended by U11.) • U11:Cancel Reservation: A Member cancels an unconcluded Reservation. (Extends U10.) • U12:Log Off: A Member logs off from iCoot. (Extends U5.) • U13:Look for CarModels: A Customer retrieves a subset of CarModels from the catalog. (Abstract, generalized by U1 and U4.)

B.2.4 Use Case Diagram The use case diagram for iCoot is shown in Figure B.11.

System Requirements

iCoot >

<

U2: View Results <<extend>>

<< inc lud e>>

U7: Make Reservation

te ex <<

nd

>>

>

>

U5: Log On

<< ex ten d> >

U10: View Reservations

tend> <<ex > d> ten ex < <

Customer

d>> <<exten

Member}

U6: View Member Details <<extend>

<<extend>>

a logged-On

NonMember U4: Search

U3: View Car Model Details {Customer is

U13: Look for Car Models

U1: Browse Index

U12: Log Off

U11: Cancel Reservation

U9: Change Password

U8: View Rentals

Member

Assistant

Figure B.11: Use case diagram for iCoot

B.2.5 Use Case Survey The use case survey for iCoot, describing how the use cases ﬁt together, is: Any Customer can look for CarModels in the catalog, by browsing the CarModel index (U1) or by searching (U4). In the latter case, the Customer speciﬁes the Categories, Makes and engine sizes that they’re interested in. Either way, after each retrieval, the Customer is shown the results as a collection of matching CarModels (U2), along with basic information such as CarModel name. The Customer can then choose to view extra information about particular CarModel objects such as a description and an advert (U3). A Customer who has become a Member can log on (U5) and gain access to extra services. The extra services are: making a Reservation (U7), canceling a Reservation (U11), checking membership details (U6), viewing outstanding Reservations (U10), changing their log-on password (U9), viewing their outstanding Rentals (U8) and logging off (U12). Assistants are involved in the life cycle of Reservations, moving Cars to and from the reserved area, for example.

461

462

Appendix B

Customers come in two varieties, Members and NonMembers. Browsing the index and searching for CarModels are two different ways of looking for CarModels (U13). In order to view CarModel details, a Customer must be viewing the results of looking for models (via the browsing or searching route). In order to reserve a CarModel, a Member must be viewing its details (NonMembers can’t make reservations, even when they’re viewing details). In order to cancel a Reservation, a Member must be viewing their outstanding Reservations.

B.2.6 Use Case Details U1:Browse Index. (Specializes U13, includes U2.)

Preconditions: None. 1. Customer selects an index heading. 2. Customer elects to view CarModels for the selected index heading. 3. Include U2. Postconditions: None. U2:View Results. (Included by U1 and U4, extended by U3.)

Preconditions: None. 1. iCoot presents Customer with a summary of each retrieved CarModel, including model number and price. 2. Extend with U3. Postconditions: None. U3:View CarModel Details. (Extends U2, extended by U7.)

Preconditions: None. 1. Customer selects one of the matching CarModels. 2. Customer requests details of the selected CarModel. 3. iCoot displays details for the selected car model (makes, engine size, price, description, advert and poster). 4. If Customer is a logged-on Member, extend with U7. Postconditions: iCoot has displayed details of selected CarModels. Non-Functional Requirements: r1. Adverts should be displayed using a streaming protocol rather than requiring a download. U4:Search. (Specializes U13, includes U2.)

Preconditions: None. 1. Customer selects required categories (if any). 2. Customer selects required Makes (if any).

System Requirements

3. Customer selects required engine sizes (if any). 4. Customer initiates the search. 5. Include U2. Postconditions: None. Abnormal paths: a1. If Customer speciﬁes no categories, makes or engine sizes, rather than retrieving the entire catalog, iCoot should not allow the search to be initiated. U5:Log On. (Extended by U6, U8, U9, U10 and U12.)

Preconditions: Member has obtained a password from their local Store. 1. Member enters the membership number. 2. Member enters the password. 3. Since iCoot must enforce one logon for a Member, Member can choose to steal (invalidate and thus take over from) an existing session. 4. Member elects to log on. 5. Extend with U6, U8, U9, U10, U12. Postconditions: Member is logged on. Abnormal Paths: a1. If the membership number/password combination is incorrect, iCoot informs Member that one of the two is incorrect (for security, they’re not told which one). a2. If the membership number/password combination is correct, but Member is already logged on and they have not elected to steal, iCoot informs Member. U6:View Member Details. (Extends U5.)

1. Member elects to view membership details. 2. Member is presented with membership details (name, address, status, amount owing, CreditCard details). 3. For security reasons, iCoot must display only the last four digits of the Member’s CreditCard number. 4. iCoot informs Member that to correct details, they must contact their local Store. Postconditions: Member has been presented with membership details. U7:Make Reservation. (Extends U3.)

Preconditions: Customer is a Member who has logged on. 1. Member elects to reserve CarModel for the details on display. 2. iCoot asks Member for conﬁrmation, issuing a warning that failure to collect a reserved CarModel will result in a ﬁne. 3. Member conﬁrms Reservation. 4. iCoot shows Member the Reservation number and indicates that Assistant will be in touch when a Car is available. 5. When an Assistant logs on to Coot, Assistant is given a list of Reservations that require action.

463

464

Appendix B

6. Assistant takes necessary action to progress Reservations (e.g. promoting to Collectable if a Car is available and moving the Car to the reserved area). Postconditions: Any requested Reservations have been made. Abnormal Paths: a1. If Member declines Reservation conditions, no Reservation is made. U8:View Rentals. (Extends U5.)

Preconditions: None. Relationships: U5. 1. Member elects to view their Rentals. 2. iCoot presents Member with summary of each Car they currently have out for rent (including number plate and due date). Postconditions: iCoot has presented Member with summaries of Cars currently rented. U9:Change Password. (Extends U5.)

Preconditions: None. 1. Member elects to change password. 2. Member enters old password (which is obscured on screen). 3. Member enters new password (obscured). 4. Member enters new password again (for conﬁrmation, also obscured). 5. Member initiates the change. 6. iCoot asks for conﬁrmation (warning that new password must be memorable). 7. If Member conﬁrms, password is changed. Postconditions: Password has been changed. Abnormal Paths: a1. If old password is incorrect or new passwords don’t match, Member is informed (but not given details of the error, for security) and password is unchanged. a2. If old passwords match but new password doesn’t follow password rules (a mix of at least six letters and digits), Member is informed and password is unchanged. U10: View Reservation objects. (Extends U5, extended by U11.)

Preconditions: None. 1. Member elects to view Reservations. 2. iCoot displays summaries of the Member’s outstanding (unconcluded) Reservations (including number, state, timestamp and CarModel number). 3. Extend with U11. Postconditions: Member has been presented with summary of outstanding Reservations. U11:Cancel Reservation. (Extends U10.)

Preconditions: None. 1. Member selects a Reservation. 2. Member elects to cancel the Reservation. 3. iCoot asks for conﬁrmation.

System Requirements

4. Member conﬁrms that they wish to cancel the Reservation. 5. iCoot marks the Reservation as Concluded and updates Assistants’ terminals accordingly. Postconditions: Any canceled Reservations that were conﬁrmed have been marked as Concluded. Abnormal Paths: a1. If Member doesn’t conﬁrm a cancellation, iCoot takes no action. U12:Log Off.

Preconditions: None. 1. Member elects to log off. 2. iCoot ends current session. 3. iCoot makes Member-only functions unavailable to Member. Postconditions: Member is logged off. Abnormal Paths: a1. For security reasons, a logged-on Member is logged off automatically if they do not interact with iCoot for ten minutes. U13:Look for CarModels (Abstract, specialized by U1 and U4.)

Preconditions: None. Postconditions: Customer has been presented with summaries of retrieved CarModels.

B.2.7 Supplementary Requirements s1. The client applet must run in Java PlugIn 1.2 (and later versions). s2. iCoot must be able to cope with a catalog of 100,000 CarModels. s3. iCoot must be able to serve 1,000,000 Customers simultaneously with no signiﬁcant degradation in performance.

B.2.8 Use Case Priorities Below is the list of use case priorities for iCoot, with the scores that were used for the ﬁrst increment. • Green: – U1: Browse Index – U4: Search – U2: View Results – U3: View CarModel Details – U5: Log On • Amber: – U12: Logoff. – U6: View Member Details – U7: Make a Reservation – U10: View Reservations

465

466

Appendix B

• Red: – U11: Cancel Reservation – U8: View Rentals – U9: Change Password During the ﬁrst increment, U6 was also completed. The other use cases were completed during the second increment.

B.3 ANALYSIS This section documents the results of the analysis phase of the iCoot development, in terms of analysis class model, a state machine for a Reservation and use case realization (communication diagrams). The Reservation state machine also applies to the full Coot system. The class model includes a few pieces from the full Coot schema, such as NonMember and dateLost.

B.3.1 Class Diagram The analysis class diagram for iCoot is shown in Figure B.12. Most of these classes also appear in the design class model (Section B.5), so their descriptions have been placed in the Glossary (Section B.8), to avoid repetition.

1..*

Car details

*

Rental

taken out by

Car Details

1

*

Car Model Details

example of Make

*

0..1

*

1

1..

rented under

Customer

1 1

makes

*

details

Car Model

*

classified as 1

*

* Reservation

sold by

*

1..

*

1..

guaranteed by

logs in with

lives at

1 1

Category

NonMember

1

Member

Vendor

1

Credit Card

Figure B.12: Analysis class diagram

Address

1

Internet Account

Analysis

B.3.2 Attributes The class attributes for iCoot are shown in Figure B.13. These attributes also appear in the design class model as ﬁelds, where they’re given types and descriptions – refer to the design documentation (Section B.5) for details.

Car

CarModel name price

travelled dateLost[0..1]

CarDetails

Member inGoodStanding: boolean number:String

Address house street county postCode

Rental number startDate dueDate totalAmount

Category name

barCode:String numberPlate vin Make name

CarModelDetails engineSize description advert poster

Vendor name

Internet Account password:String

CreditCard number:String type:String expiryDate

Customer name phoneNumber: String amountDue:int

Reservation number:String timestamp state

NonMember driversLicense: String

Figure B.13: Analysis attributes

B.3.3 Operation List • CarModel: – getSummary() – Fetch a summary of the receiver, including model number and price. – getDetails() – Fetch the receiver’s details, including makes, engine size, price, description, advert and poster. • CarModelHome: – ﬁndByIndexHeading(h:String) – Search for CarModel objects under index heading h. – ﬁndByQuery(categories,makes,sizes) – Search for CarModel objects with Category from categories, a Make in makes and engine size in sizes. • LogonController: – logon(n:String,p:String,s:boolean) – Log on the Member with membership number n and password p, specifying whether or not to steal any existing session with s.

467

468

Appendix B

•

• •

•

• •

– changePassword(m:Member,o:String,n1:String,n2:String) – Change the password for m to n1, as long as n2 matches and the current password is o. – logoff() – Log off the logged-on Member. Member: – getPassword():String – Fetch the receiver’s password. – isLoggedOn():boolean – True if the receiver is logged on. – logon() – Log the receiver on. – logoff() – Log the receiver off. – getDetails() – Fetch the receiver’s details, including name, address, status, amount owing and (concealed) credit card details – setPassword(p:String) – Set the receiver’s password to p. MemberHome: ﬁndByMembershipNumber(n:String):Member – Find the Member with membership number m. MemberUI: – search(categories,makes,sizes) – Search for CarModel objects with a Category from categories, a Make in makes and engine size in sizes. – index(h:String) – Search for CarModel objects under index heading h. – logon(n:String,p:String,s:boolean) – Log on the Member with membership number n, password p, specifying whether or not to steal any existing session with s. – setMember(m:Member) – Set the logged-on Member to m. – showMemberDetails() – Show details for the logged-on Member. – showRentals() – Show Rental objects for the logged-on Member. – showReservations() – Show unconcluded Reservation objects for the logged-on Member. – changePassword(o:String,n1:String,n2:String) – Change the password for the logged on Member to n1, as long as n2 matches and the current password is o. – conﬁrmChange() – Conﬁrm that the password really should be changed. – reserve(c:CarModel) – Reserve c for the logged-on Member. – conﬁrmReserve() – Conﬁrm that the Reservation really should be made. – cancel(r:Reservation) – Cancel r. – conﬁrmCancel() – Conﬁrm that the Reservation really should be canceled. – showDetails(c:CarModel) – Show details for c. – logoff() – Log off the logged-on Member. NonMemberUI: – search(categories,makes,sizes) – Search for CarModel objects with a Category from categories with a Make in makes and engine size in sizes. – index(h:String) – Search for CarModel objects under index heading h. Rental: getSummary() – Fetch a summary of the receiver, including number plate and due date. RentalHome: ﬁndByMember(m:Member – Fetch the Rental objects for member m.

Analysis • Reservation: – getSummary() – Fetch a summary of the receiver, including number, timestamp, state and CarModel. – getNumber() – Fetch the receiver’s number. – setState(s) – Set the receiver’s state to s. • ReservationHome: – ﬁndByMember(m:Member) – Fetch the reservations for m. – create(c:CarModel,m:Member) – Reserve c for m, with the current date and time.

B.3.4 State Machine for a Reservation Figure B.14 shows the state machine diagram for a Reservation, produced to model its complex life cycle. The accompanying state machine survey is: no car found no car arrives

Waiting

member cancels

car assistant returned finds car

member renews

Notifiable

member unreachable member cancels

assistant notifies member member cancels

Needing Renewal

member no-show

Collectable member collects

member cancels

member cancels

Storable

member unreachable

assistant stores

Concluded

Figure B.14: State machine diagram for a Reservation

When a Member reserves a CarModel over the Internet, the Reservation is initially Waiting to be processed by an Assistant (this is so the Customer can make a Reservation without the intervention of an Assistant). The Reservation becomes Notiﬁable if, some time later, an Assistant ﬁnds a suitable unreserved Car in the display area of the car park, or if one is returned by a Customer. In this case, the Car is moved to the reserved area. If no Car becomes available for a particular Reservation within a week, the Reservation becomes NeedingRenewal: the Member must be contacted, by phone or in person, so

469

470

Appendix B

that they can cancel the Reservation, or ask for it to be renewed for another week. If the Member cancels or can’t be contacted within ﬁve days, the Reservation is Concluded. Once a Reservation is Notiﬁable, the Member must be notiﬁed by an Assistant, in person or by phone, within three days; if the Customer can be reached, the Reservation is Collectable otherwise it becomes Displayable (a Car that was moved to the reserved area must be returned to the display area). Once a Reservation is Collectable, the Member must collect the Car within three days: if they do collect, the Reservation is Concluded; otherwise, the Reservation becomes Displayable. Once a Displayable Reservation’s Car has been put back in the display area, the Reservation is Concluded. At any time, the Member may cancel the Reservation over the Internet, by phone or in person. The system will keep Assistants informed as to the state of current (not yet concluded) reservations, so that they can take appropriate action.

B.3.5 Use Case Realization The communication diagrams for iCoot, verifying the analysis class model, are shown in Figures B.15 through B.26, one per system use case. Note the use of guards (arbitrary conditions in brackets), to specify conditional messages and * to specify iteration (iteration guards can be used to control iteration, but these would have made the diagrams more complex).

1:index(indexHeading)

Customer

)

ing

d ea

:NonMemberUI or :MemberUI

nd

g(i

in ad

exH

e

find

st= 1:li

1.

d yIn

exH

B

1.2: Include U2, passing in list.

:CarModel Home

Figure B.15: Communication diagram for U1:Browse Index

Analysis

2:carModels(summaries) Customer

:NonMemberUI or :MemberUI

1*: um etS

s=g ) ry(

ma

For each CarModel in list (from including use case), add s to summaries. Each summary includes model number and

:CarModel

price.

Figure B.16: Communication diagram for U2:View Results

1:showDetails(c) Member

1.2:carModelDetails(d) :MemberUI

)

ls(

i eta

tD

ge

:d= 1.1

Details include makes, engine size, price, description, advert and poster.

c:CarModel

Figure B.17: Communication diagram for U3:View CarModel Details

471

472

Appendix B

Parameters: c is categories m is makes e is engine sizes

1:search(c,m,e) Customer

:NonMemberUI or :MemberUI

,e)

c,m

1.1

:

=f list

ind

By

Q

ry( ue

1.2: Include U2, passing in

:CarModel Home

list

Figure B.18: Communication diagram for U4:Search

Member

1:

lo

go

n(

n,

1.

p, s

7:

)

gr e

et

() 1.1:logon

(n,p,s)

1.6:setMem

:MemberUI

ber(m)

)

b

:Logon Controller

pN

em

hi ers

b

yM

B nd =fi

:m

1.2

:Member Home

Parameters: n is membership number p is password s is steal

() ord sw as n() etP dO =g ge :p2 og ull] isL n-n :o= ) no p)] [m ls( gon( 1.3 ua o .eq t o]:l o [p2 1.4 s or n [ 1.5

um

n er(

m:Member

Figure B.19: Communication diagram for U5:Log On

Analysis

1:showMemberDetails() 1.2:memberDetails(d)

Member

:MemberUI

)

ls(

1.1

:d=

g

e etD

tai

Details include name, address, status, amount owing and censored CreditCard details.

:Member

Figure B.20: Communication diagram for U6:View Member Details

m:Member c:CarModel

1:reserve(c) 1.1:fineWarning() 1.1.1:confirmReserve() 1.1.1.3:reservationNumber(n)

Member

r()

be

m

Nu

t ge

.

1.

1 1.

2:

n=

:MemberUI

1.

1.

1.

1:

r=

cr

ea

te

(c ,m

)

r:Reservation

:Reservation Home

Figure B.21: Communication diagram for U7:Make Reservation

473

Appendix B

1:showRentals() 1.3:rentals(summaries) Member

:MemberUI

(m)

ary(

)

e

yM

B find

er mb

umm

ist= .1:l

:s=g

etS

1

1.2*

474

:Rental Home For each Rental in list, add s to summaries.

m:Member

Each summary

:Rental

includes number plate and due date.

Figure B.22: Communication diagram for U8:View Rentals 1:c

ha

Member

ng e 1.1 Pas :m swo em r 1.1 or d(o,n .1: ab co i lty 1,n2 nfi W rm ar ) Ch nin an g( ge ) ()

Parameters: o is old password n1 is new password n2 is new password again

1.1

.1.

:MemberUI

1:c

ha

ng

eP

as

sw

ord

(m

,o,

n1

1.1.1.3

[p.equ

a

,n2

ord() (n1) sword tPassw etPas :p=ge :s .2 )] 2 .1 1.1 ls(n .equa nd n1 ls(o) a

)

:Logon Controller

m:Member

Figure B.23: Communication diagram for U9:Change Password

Analysis

1:showReservations()

1.3:reservations(summaries)

Member

:MemberUI

)

(m ber

mar

y()

em

yM

B find list=

1.2* :s

=ge

tSu m

:

1.1

:Reservation Home

m:Member For each Reservation in list, add s to summaries. Each summary

:Reservation

includes number, timestamp, state, car model.

Figure B.24: Communication diagram for U10: View Reservations 1:cancel(r)

1.1:confirmationNeeded() 1.2:confirmCancel() Member

)

:MemberUI

d de

clu

n co

(

te

t

se

: .1

a St

2

1.

r:Reservation

Figure B.25: Communication diagram for U11:Cancel Reservation

475

476

Appendix B

Member

1:

lo

go

ff( )

1.1:logoff (m)

:MemberUI

)

off(

og .2:l

1

:Logon Controller

m:Member

Figure B.26: Communication diagram for U12:Log Off

B.4 SYSTEM DESIGN This section documents the results of system design carried out during the design phase of the iCoot development, in terms of technology choices, layers, packages, deployment diagram, security policy and concurrency policy.

B.4.1 Technology Choices On the client side, the choice of technology is driven by convenience for the customer – we do not want customers to have to install any software in order to access our services. Also, we want them to be able to use any desktop machine, regardless of the operating system they have installed. The obvious choice for the client environment, therefore, is a web browser. Since the user interface must be interactive, in order to make a Reservation for example, we have to choose between technologies such as HTML/CGI, Java applets, ActiveX controls and Flash. Due to the need for portability (and client security), we can discount ActiveX controls. We would also prefer our customers to have almost-instant access to the user interface when browsing our site. This effectively discounts applets and Flash, both of which typically

System Design

involve a delay while the interface is downloaded. Therefore, the initial user interface will be HTML/CGI. Once on the server side, servlets are a good choice for processing CGI requests because they’re portable, they’re efﬁcient and they have access to all the facilities of J2EE, which provides everything that the servlets might need (such as access to distributed transaction management). Traditional scripting raises issues with portability, expressive power and performance; .Net technologies raise issues with portability. Once servlets are being used as the entry point to our servers, the obvious choice for producing dynamic web pages is the JSP mechanism. Initially, we will use an open-source (free) implementation of J2EE for development and deployment, being careful to avoid any proprietary lock-ins. The implementation must support the forwarding of requests to servlets and JSPs running in a separate process, so that the latter can be accessed directly by GUI clients. Should the open-source implementations prove inadequate, we can simply purchase a commercial product and redeploy our code. Because of their portability, we can deploy our servlets on any combination of hardware and operating system, and then redeploy at a later date if necessary. Initially, each store will have the system software deployed on two budget Linux servers, providing fail-over and throughput without the expense of a higher-powered server or highly-available hardware. For the business data, we will use an open-source database initially, switching to a commercial product later if necessary. A relational database will be used because of the maturity of the technology and because our application is business-oriented, with a large quantity of data but no great logical complexity. The database will also be deployed on a pair of Linux servers at each store. For the future, we aim to provide a user interface that can be used on a mobile phone or PDA. We would prefer to avoid WAP on devices that can’t manage the HTML/CGI interface because it tends to be clumsy and unpopular. Instead, we will use J2ME, allowing us to provide a rich interactive experience that scales automatically to the size of the screen. For those customers who wish to install J2SE on their client, either manually or by using their Web browser’s plug-in mechanism, we will also provide a conventional graphical user interface as an applet. This applet will be deployed on touch-screen kiosks in each store to enhance the customer experience, when viewing adverts for example. It is anticipated that the J2ME and J2SE interfaces will bypass the JSPs and servlets, for improved performance.

B.4.2 Layer Diagram iCoot layers are illustrated in Figure B.27. Persistence is provided by the JDBCLayer, using the standard JDBC library to access a relational database. There is no separate persistence layer, because we expect a relational database to serve our needs for the lifetime of the system.

477

478

Appendix B

HTMLLayer

SwingLayer

MicroLayer

HTTPCGILayer

ControlLayer

ServletsLayer

RMILayer

ServerLayer BusinessLayer

JDBCLayer

Figure B.27: iCoot layer diagram

The BusinessLayer contains implementations of the entity objects from the analysis class diagram, along with various supporting objects. These objects contain JDBC code for shipping data to and from the database. The ServerLayer translates the objects and messages in the BusinessLayer into business services, in the form of messages on server objects. Objects in the ServerLayer are EJB session objects, which have two beneﬁts: ﬁrstly, they give us access to J2EE transaction management; secondly, they allow us to provide direct access for GUI clients over RMI, bypassing the WebServer. In order to keep the ServerLayer closed, all information returned by the business services takes the form of protocol objects, lightweight copies of the business objects. The ServletsLayer is a control layer for HTML/CGI clients. Each servlet translates one or more objects in the ServerLayer into simple commands and questions that can be issued from the client. In response to each command or question, a servlet will perform whatever actions are necessary and then pass the next HTML page back to the client. So that page design and source code are separate, every reply page is built by a JSP that produces content dynamically based on the customer’s interaction. The JSPs receive their dynamic data as protocol objects passed in by the servlets. The network communication for the HTMLLayer is provided by the standard HTTPCGILayer. The RMILayer is a network layer allowing remote access from GUIs (Java applications and any device using J2ME). The objects in this layer are simply decorators for the EJB session objects in the ServerLayer: each server object is decorated with an RMI servant, while each RMI servant is accessed via an RMI proxy on the client. When communicating with the ControlLayer, the RMILayer uses the same protocol objects that the ServletsLayer uses when invoking JSPs.

System Design

The ControlLayer sits between the GUI objects and the RMI proxies. It serves to simplify interaction with the server objects and to hide the details of RMI. The RMILayer, ControlLayer, SwingLayer and MicroLayer are not documented fully because the graphical user interfaces are not part of the ﬁrst increment of iCoot.

B.4.3 Layer Interaction Policy On the server, for the sake of simplicity, all layer communication will ﬂow downwards. In other words messages will only be sent from one layer to the layer below. Events will be used on the client side for the beneﬁt of the SwingLayer and the MicroLayer, so that application-speciﬁc knowledge can be pushed from the user interface components down to the ControlLayer. (The HTML/CGI front end does not need events because all information displayed to the user is calculated by the servlets and passed directly to JSPs for presentation.) Layers will be closed, in order to make implementation and maintenance easier: each object will be able to access objects in the layer immediately below, but not beyond.

B.4.4 Packages The package diagram for Coot (including the graphical user interface/RMI packages not implemented in the ﬁrst increment) is shown in Figure B.28.

B.4.5 Deployment Diagram The deployment diagram for iCoot is shown in Figure B.29. The iCoot data tier comprises two database servers (which we have called DBServer). Having two such nodes improves throughput and reliability. Each DBServer hosts a DBMS process for managing access to data. The cootschema.ddl artifact contains commands for creating database tables, in a format speciﬁc to the database being used. This is deployed to each DBMS process, using database-speciﬁc tools (no detail given here). Note that cootschema.ddl contains the schema for the full Coot system, since iCoot and Coot use the same data. The middle tier, which communicates with the data tier, consists of two server machines (CootServer), again duplicated for the sake of reliability and throughput. Each CootServer hosts a CootBusinessServer (for handling business requests) and a WebServer (for handling static HTML content and forwarding business requests to the CootBusinessServer). Data access for the CootBusinessServer is provided by the DBMS. Because they’re proprietary to the products that we select, the communication protocols between the WebServer

479

480

Appendix B

com::nowhere

swing

micro

control

protocol

rmi

servlets

server

business

Figure B.28: iCoot package diagram

and the CootBusinessServer and between the CootBusinessServer and the DBMS are not speciﬁed. Within each CootServer, the iCoot folder, containing static HTML pages, is deployed to the WebServer, while the icoot.ear archive is deployed to the CootBusinessServer. The icoot.ear archive contains servlets, JSPs, business objects and (eventually) RMI decorators, from the com::nowhere package.

System Design

<<device>> CootHTMLClient

*

* <<device>> CootGUI Client

<<ExecutionEnvironment>> WebBrowser

HTTP

JRMP 2

<<device>> CootServer

<<ExecutionEnvironment>> WebServer

<<ExecutionEnvironment>> CootBusinessServer

<<use>> icoot.ear

iCoot

<<manifest>> com::nowhere servlets protocol server business

2

<<device>> DBServer

<<ExecutionEnvironment>> DBMS

cootschema.ddl

Figure B.29: iCoot deployment diagram

481

482

Appendix B

Each CootServer can be accessed simultaneously by any number of CootHTMLClients. Each hosts a WebBrowser, which accesses one of the WebServers using HTTP. No artifacts need to be deployed to the CootHTMLClients. Eventually, we will also provide access from CootGUIClients. Each CootGUIClient will access one of the CootServers, using JRMP. Because the mechanism that allows such requests to get into the CootBusinessServer is the subject of a future increment, no details are given. Nor is any detail given for the CootGUIClient processes. The artifacts deployed to the CootGUIClients, if any, are not speciﬁed. CootHTMLClient

B.4.6 Security Policy Searching and browsing services will be available to all-comers, without logging in. In contrast, for Member-only services, each Member must ﬁrst obtain a password in person from their local Store, then use it to log in to the Member’s area from their chosen client. The membership numbers and passwords used to log in will be managed in a central directory, for ease of maintenance, using the standard Java integration mechanisms. In order to keep Member activity private, all access to Member services from the client will be over SSL rather than plain TCP/IP. SSL will also be employed between servers for internal protection. The servers will be deployed behind an Internet ﬁrewall so that external access can be tightly controlled.

B.4.7 Concurrency Policy Objects in the BusinessLayer will be managed using distributed transactions. At the start of each business request (i.e. each method on a ServerLayer object), a Java transaction will be created – this transaction will be associated with every database access made by business objects within that request. At the end of each request, the Java transaction will be committed, thus making updates available to other requests. In order to minimize transaction conﬂict, all RMI servants, servlets and server objects will be stateless. For GUI clients, access to local data (copies of protocol objects) will be single-threaded. For HTML clients, each JSP will have exclusive access to its protocol objects, so such access is also effectively single-threaded. For business data, low-level concurrency control is managed automatically by the EJB framework: every use of a business service is wrapped inside a transaction, which passes right through to the database management system.

Subsystem Design

To simplify concurrency control at the business level, two strategies will be employed: ﬁrstly, single log-on will be enforced for Members. Secondly, updates to the Catalog of available CarModels will be made off-line and switched with the live Catalog in the early hours of the morning. This will minimize the need to report errors such as ‘Attempt to reserve a car model that has been withdrawn’ to Customers and Assistants. (This will still occasionally happen, because client displays won’t be updated explicitly when data changes on the server via concurrent paths; pushing all relevant changes to clients, although technically feasible, would be too inefﬁcient.)

B.5 SUBSYSTEM DESIGN This section documents the results of subsystem design, carried out during the design phase of the iCoot development, in terms of business services, design class models (one per layer plus the protocol objects), database schema, user interface design and business service realization (sequence diagrams). In general, this documentation describes the subsystem artifacts required to support business service realization and the CootHTMLClient.

B.5.1 Business Services 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13.

Read headings from the CarModel index. Read CarModels for a given index heading. Read all CarModel Categories. Read all CarModel engine sizes. Read all Makes of CarModel. Read CarModels for a given set of Categories, engine sizes and Makes. Read details for a given CarModel. Reserve a CarModel. Read details for a given Member. Change a Member’s password. Read the Cars rented by a given Member. Read the Reservations made by a given Member. Cancel a Reservation.

B.5.2 ServletsLayer Class Diagram The class diagram for the ServletsLayer is shown in Figure B.30.

483

484

Appendix B

pkg servlets

Authentication Servlet

Reservations Servlet

Membership Servlet

Rentals Servlet

Catalog Servlet

Figure B.30: ServletsLayer classes

B.5.3 ServletsLayer Field List All servlets are stateless and therefore do not have any ﬁelds. Any information about the state of a customer’s interaction is recorded in the result pages themselves and in a state object stored in the browser’s HTTP session. For logged-on members, the session identiﬁer is also stored in the HTTP session. Objects in the server layer are located using their homes (implemented using the Singleton pattern).

B.5.4 ServletsLayer Message List Using the standard Java mechanism, each client request is passed to the selected servlet via a message called doGet(:HttpServletRequest,:HttpServletResponse). The questions and commands from the client are passed in as part of the HttpServletRequest. Below is a list of the questions and commands that can be passed in by the client, along with the JSPs that are invoked as a result. AuthenticationServlet

• logon From home page; takes a membership number, password and ‘steal’ parameter and, upon successful logon, returns the member page. (The steal parameter indicates whether the member would like to steal an existing session.) • logoff From member page; logs off the current Member.

Subsystem Design

CatalogServlet

• index From member or non-member page; returns the index page containing index headings to choose from. • browse From index page; takes an index heading as parameter and returns the results page containing matching CarModels. • search From member or non-member page; returns the search page containing every Category, Make and engine size to choose from. • query From search page; takes Category ids, Make ids and engine sizes as parameters and returns the results page containing matching CarModels. • details From results page; takes a CarModel id as parameter and returns the details page containing details for that CarModel. MembershipServlet

• membership From member page; returns the membership page, containing details for the current Member. • password From member page; returns the password page. • changePassword From member page; takes an old password and a new password as parameters and returns the conﬁrmChange page. • conﬁrmChange From the conﬁrmChange page; sets the new password for the current Member, if the old password given was correct. RentalsServlet

• rentals

From member page; returns the rentals page, containing Rentals for the current

Member. ReservationsServlet

• reserve From details page; takes a CarModel id as parameter and returns the conﬁrmReserve page. • conﬁrmReserve From conﬁrmReserve page; reserves the CarModel already identiﬁed and returns the conﬁrmation page. • ok From conﬁrmation page; returns the details page. • reservations From member page; returns the reservations page, containing Reservations for the current Member. • cancel From the reservations page; takes the Reservation id as parameter and returns the conﬁrmCancel page. • conﬁrmCancel From conﬁrmCancel page; cancels the previously selected Reservation and returns the reservations page.

B.5.5 ServerLayer Class Diagram The class diagram for the ServerLayer is shown in Figure B.31. Each class also has a home, implemented using the Singleton pattern (not shown).

485

486

Appendix B

pkg server

Authentication Server

Rentals Server

Rentals ServerHome

Authentication ServerHome

Catalog Server

Catalog ServerHome

Membership Server

Membership ServerHome

Reservations Server

Reservations ServerHome

Figure B.31: ServerLayer classes

B.5.6 ServerLayer Field List The server objects are stateless, thus they have no ﬁelds. (All BusinessLayer classes are accessed via their respective homes.)

B.5.7 ServerLayer Message List For the server objects, the messages below correspond to the business services. (Homes have been omitted from this list because they simply create server objects, with no parameters.) AuthenticationServer

• +logon(n:String,p:String,s:boolean):long Log on the Member with membership number n and password p, specifying whether or not to steal any existing session with s. • +logoff(i:int) Log off the Member with session identiﬁer i.

Subsystem Design

CatalogServer

• • • •

Read the names of every Category. +readMakeNames():String[] Read the names of all Makes. +readEngineSizes():int[] Read the unique engine sizes of all CarModel details. +readIndexHeadings():String[] Read the index headings, derived from all CarModel numbers and Make names. • +readCarModels(h:String):PCarModel[] Read all CarModels matching the index heading h. • +readCarModelDetails(i:int):PCarModelDetails Read details for the CarModel with identiﬁer i. • +readCarModels(q:PCatalogQuery):PCarModel[] Read all CarModels that match the query q. +readCategoryNames():String[]

MembershipServer

• +readMember(i:int):PMember Read the Member with session identiﬁer i. • +changePassword(i:int,o:String,n:String) Change the password for the Member with session identiﬁer i, using old password o and new password n.

RentalsServer

• +readRentals(i:int):PRental[]

Read all Rentals for the Member with session identiﬁer i.

ReservationsServer

• +readReservations(i:int):PReservation[] Read all Reservations for the Member with session identiﬁer i. • +createReservation(i:int,c:int) Create a Reservation for the Member with session identiﬁer i and the CarModel with identiﬁer c. • +cancelReservation(i:int,r:int) Cancel the Reservation with identiﬁer r, for the Member with session identiﬁer i, as long as the Reservation matches the Member.

B.5.8 BusinessLayer Class Diagram The class diagram for the BusinessLayer is shown in Figure B.32. Most of these classes also appear in the analysis class model (Section B.3), so their descriptions have been placed in the Glossary (Section B.8), to avoid repetition. Each entity class has a home, apart from Customer (because it’s abstract). ReservationStateHome is an Abstract Factory for creating instances of its subclasses. The Store class, used for illustration in Chapter 13, has not been included in this appendix since it plays no part in business service realization.

487

488

Appendix B

pkg business

Car Details

-cars:List

-details Car 1

Rental

0..1

1..*

Address

*

*

-address -customer 1

1 1..*

Member

Customer

1 -customer 1

-model CarModel Details

-details

*

1

1 *

*

-account

-card

-model

CarModel

1..*

Reservation *

1

Internet Account

Credit Card

1

*

1

-vendor 1 Vendor

1

-category

1..*

Category

Waiting

Needing Renewal

Reservation StateHome

1

-makes:List

Catalog Query

ReservationState

Make

Notifiable

InternetAccount Home

-state

Collectable

Member Home

Car Home

Rental Home

CreditCard Home

Category Home

Vendor Home

Address Home

Displayable

CarModel Home

Reservation Home

CarModel DetailsHome

Concluded

Catalog Query Home

Make Home

CarDetails Home

Figure B.32: BusinessLayer classes

B.5.9 BusinessLayer Field List The list below shows ﬁelds for stored attributes only; ﬁelds that store links are shown in Figure B.32. With the exception of the state objects, every object has an -id:int that stores the universal identiﬁer but which has been omitted from the list. Classes with no ﬁelds other than links have been omitted.

Subsystem Design

Address

• • • •

House number and/or name (unique within the postal code). -street:String Street in which the house stands. -county:String County where the street is. -postCode:String Postal code for the sorting ofﬁce and region. -house:String

Car

• -traveled:int The distance, in kilometers, that the Car has traveled (taken from the odometer). • -dateLost:Date Date the Car was reported missing, null if not lost. CarDetails

• -barCode:String The bar code, attached inside the Car’s windscreen. • -numberPlate:String The Car’s license number as it appears on the plate. • -vin:String The Car’s unique Vehicle Identiﬁcation Number, on a plate riveted to the body. CarModel

• -name:String The Make’s unique name for the CarModel. • -price:int The daily cost of hiring the Car, in cents. CarModelDetails

• • • •

The capacity of the engine, in cubic centimeters. A single-sentence description of the CarModel. -advert:String The ﬁle name of a streaming advert for the CarModel. -poster:String The ﬁle name of a poster for the CarModel. -engineSize:int

-description:String

CatalogQuery

• -makeIds:List Universal identiﬁers for relevant Makes (using identiﬁers avoids retrieving Makes when building the database query). • -carModelIds:List Universal identiﬁers for relevant CarModels (using identiﬁers avoids retrieving CarModels). • -engineSizes:List Engine sizes, in cubic centimeters. Category

• -name:String

The name of the Category.

Collectable

• -dateNotiﬁed:Date

Date the Customer was notiﬁed that the car is ready for collection.

489

490

Appendix B

Concluded

• -reason:String Why the Reservation is concluded: ‘‘Successfully rented’’, ‘‘Canceled by customer’’, ‘‘Wasn’t renewed’’ or ‘‘Wasn’t collected’’. CreditCard

• -type:String Type of Card (e.g. ‘‘Annex’’). • -number:String Number on the Card. • -expiryDate:Date Date Card expires. Customer

• -name:String The Customer’s name. • -phone:String The Customer’s phone number. • -amountDue:int Fees due, in cents, for the Customer (for example, for overdue Rentals). Displayable

• -reason:String Why the Car must be put back in the display area, either ‘‘Customer uncontactable’’ or ‘‘Failed to collect’’. InternetAccount

• -password:String Password for the associated Member; must be a mix of at least six letters and digits. • -sessionId:long Session identiﬁer for the associated Member: 0 if not logged on; securely random, unique, non-zero number if logged on. Make

• -name:String

The name of the manufacturer.

Member

• -number:String The Member’s unique number. • -inGoodStanding:boolean Whether or not the Member has any outstanding issues, such as disputed late fees. NeedingRenewal

• -dateRenewalNeeded:Date Date by which the Reservation must be renewed to avoid becoming automatically concluded. Notiﬁable

• -datePutAside:Date Date the Car was moved to the reserved area.

Subsystem Design

Rental

• • • •

The unique number for the Rental. -startDate:Date The date the Rental was taken out. -dueDate:Date The date the Rental is due to end. -totalAmount:int The amount paid for the Rental, in cents. -number:String

Reservation

• -number:String The unique number for the Reservation. • -timestamp:Timestamp The date and time that the Reservation was made. Vendor

• -name:String

Name of the Vendor.

Waiting

• -lastRenewedDate:Date date it was created).

Date the Reservation was last renewed (initially the same as the

BusinessLayer Message List Most of the public messages on the BusinessLayer classes are simply accessors for attributes, derived attributes or links, so no detail is given for these. For the ReservationState hierarchy, and the Reservation class itself, every class has one message for each event shown in Figure B.14 and one getter for each state attribute. Details of these messages have been omitted, for brevity. In addition, the Reservation class has test messages to enable clients to discover which state the receiver is in, e.g. isConcluded and isWaiting. For the homes, details resulting from the Singleton pattern are not given. To support the homes, every class created by a home has a package constructor that takes all attributes and links as parameters; no further detail is given for these constructors. Apart from ReservationStateHome and CatalogQueryHome, every home has the following messages (where X is a stand-in for the corresponding BusinessLayer class): • +ﬁndByPrimaryKey(id:int):X Returns the instance of X with universal identiﬁer id. • +create(...):X Takes every attribute and link as parameters and returns a new instance of X. ReservationStateHome has one creation message for each subclass, taking all subclass attributes

as parameters. has a create message that takes three List parameters: makeIds, categoryIds and engineSizes. CatalogQueryHome

491

492

Appendix B

All other home messages used in business service realization are listed below. CarModelHome

• +ﬁndByIndexHeading(h:String):List Returns all CarModels that have h as a model number or Make name, sorted by model name. • +ﬁndByQuery(q:CatalogQuery):List Returns all CarModels that have a Make, a Category and an engine size that appear in q, sorted by model name. CarModelDetailsHome

• +ﬁndByCarModelId(id:int):CarModelDetails Returns the instance that matches the CarModel with universal identiﬁer id. • +ﬁndEngineSizes():List Returns all engine sizes, in ascending order. CategoryHome

• +ﬁndCategoryNames():List<String> Returns all Category names, in alphabetical order. MakeHome

• +ﬁndMakeNames():List<String[]>

Returns all Make names, in alphabetical order.

MemberHome

• +ﬁndByMembershipNumber(n:String):Member Returns the Member with membership number n. • +ﬁndBySessionId(id:long):Member Returns the Member with session identiﬁer id. RentalHome

• +ﬁndByMember(m:Member):List

Returns all Rentals for m, sorted by start date.

ReservationHome

• +ﬁndUnconcludedByMember(m:Member):List vations for m, sorted by creation date.

Returns all the unconcluded Reser-

B.5.10 Protocol Objects Class Diagram The class diagram for the protocol objects, used for communication between the server layer and the servlets layer, is shown in Figure B.33.

Protocol Objects Field List The list below shows ﬁelds for stored attributes only; ﬁelds that store links are shown in Figure B.33. The meaning of these ﬁelds is the same as those for the BusinessLayer classes, with the exception that cardNumber only includes the last four real digits. The message variable on PServerException stores an explanation for the exception.

Subsystem Design

pkg protocol

PCar -cars:List 1

PAddress

*

-details PCar Details

-address

1

1

1 1

PRental PMember

PServer Exception

1 -card

1

-model PCarModel

PReservation 1

1

-category

PCarDetails Home

PCard Home

1

1

1..*

-makes:List

PCategory

PMake

PCar Home

PServer Exception Home

PCategory Home

PCard

1

PCatalog Query

PReservation Home

PCarModel Home

PCarModel Details

PRental Home

PMake Home

PMember Home

PCatalog QueryHome

PAddress Home

PCarModel DetailsHome

Figure B.33: Protocol classes

• • • • • • • • • •

PAddress -house:String, -street:String, -county:String, -postCode:String PCar -traveled:int, -numberPlate:String PCarModel -id:int, -name:String, -price:int PCarModelDetails -engineSize:int, -description:String, -advert:String, -poster:String PCatalogQuery -makeIds:int[], -carModelIds:int[], -engineSizes:int[] PCategory -name:String PCreditCard -type:String, -number:String PMake -name:String PMember -name:String, -phone:String, -amountDue:int, -inGoodStanding:boolean PRental -startDate:Date, -dueDate:Date

493

494

Appendix B

• PReservation -id:int, -number:String, -timestamp:Timestamp • PServerException -message:String

Protocol Objects Message List Nearly all the messages on the protocol classes are accessors for attributes or links. In addition, each protocol class has a toString message that returns a human-readable summary of the receiver: these messages are used to display objects in user interfaces. Each home has a single create message with a parameter to initialize each attribute and link. Beyond this, all messages on homes are derived from the Singleton pattern. To support the homes, each protocol class has a package constructor that takes every attribute and link as parameters. No further detail is given for these constructors. In addition to its basic creation message, PCatalogQueryHome has a createCatalogQuery message that takes three int[] parameters (makeIds, categoryIds and engineSizes) and returns a new CatalogQuery. Since none of the protocol classes or homes have any special messages, no detailed list is given here.

B.5.11 Database Schema The database schema is shown in Figure B.34. In this diagram, the names of primary key columns are shown in bold and the names of foreign key columns are shown in italics. The meaning of the attribute columns is the same as those for the ﬁelds in the BusinessLayer classes. Of all the columns, only DATELOST in the CAR table is nullable.

B.5.12 User Interface Design The user interface design is not given here since it is similar to the user interface sketches shown in Section B.3.

B.5.13 Business Service Realization Figures B.35 through B.46 show one sequence diagram, documenting business service realization, per system use case. Several diagrams use frames to show loops. The loop operator has a pseudocode guard controlling the number of iterations. Frames have also been used to enclose lifelines that refer to included use cases, using the ref operator – in this case, the name of the included use appears in the center of the frame, alongside a list of the parameters that are passed in. For the sake of simplicity, each interaction between a customer’s browser and a servlet has been shown as a message sent from the actor directly to the servlet. In reality, this is implemented by passing a command and its parameters into the servlet’s doGet message.

Subsystem Design

ADDRESS (ID:INTEGER,HOUSE:VARCHAR(99),STREET:VARCHAR(99),COUNTY:VARCHAR(99), POSTCODE:VARCHAR(99)) CAR (ID:INTEGER,TRAVELLED:INTEGER,DATELOST:DATE,CARDETAILSID:INTEGER) CARD (ID:INTEGER,TYPE:VARCHAR(99),NUMBER:VARCHAR(99)) CARDETAILS (ID:INTEGER,BARCODE:VARCHAR(99),NUMBERPLATE:VARCHAR(99),VIN:VARCHAR(99)) CARMODEL (ID:INTEGER,NAME:VARCHAR(99),PRICE:INTEGER,CARMODELDETAILSID:INTEGER, CATEGORYID:INTEGER,VENDORID:INTEGER) CARMODELDETAILS (ID:INTEGER,ENGINESIZE:VARCHAR(99),DESCRIPTION:VARCHAR(256), ADVERT:VARCHAR(99),POSTER:VARCHAR(99)) CATEGORY (ID:INTEGER,NAME:VARCHAR(99)) COLLECTABLERESERVATION (RESERVATIONID:INTEGER,DATENOTIFIED:DATE) CONCLUDEDRESERVATION (RESERVATIONID:INTEGER,REASON:VARCHAR(99)) CUSTOMER (ID:INTEGER,NAME:VARCHAR(99),PHONE:VARCHAR(99),AMOUNTDUE:INTEGER) DISPLAYABLERESERVATION (RESERVATIONID:INTEGER,REASON:VARCHAR(99)) INTERNETACCOUNT (ID:INTEGER,PASSWORD:VARCHAR(99),SESSIONID:INTEGER) MAKE (ID:INTEGER,NAME:VARCHAR(99)) MAKECARMODEL (CARMODELID:INTEGER,MAKEID:INTEGER) MEMBER (ID:INTEGER,NUMBER:VARCHAR(99),INGOODSTANDING:BOOLEAN,CARDID:INTEGER, ADDRESSID:INTEGER) NEEDINGRENEWALRESERVATION (RESERVATIONID:INTEGER,DATERENEWALNEEDED:DATE) NONMEMBER (ID:INTEGER,DRIVERSLICENSE:VARCHAR(99)) NOTIFIABLERESERVATION (RESERVATIONID:INTEGER,DATEPUTASIDE:DATE) RENTAL (ID:INTEGER,NUMBER:VARCHAR(99),STARTDATE:DATE,DUEDATE:DATE,TOTALAMOUNT:INTEGER) RENTALCAR (RENTALID:INTEGER,CARID:INTEGER) RESERVATION (ID:INTEGER,NUMBER:VARCHAR(99),TIMESTAMP:TIMESTAMP,CUSTOMERID:INTEGER, CARMODELID:INTEGER) VENDOR (ID:INTEGER,NAME:VARCHAR(99)) WAITINGRESERVATION (RESERVATIONID:INTEGER,LASTRENEWEDDDATE:DATE)

Figure B.34: Database schema

495

ref

U2(list)

readCarModels("Kya")

String[]

:CarModel Home

List

list=findByIndexHeading("Kya")

Figure B.35: Sequence diagram for U1:Browse Index

browse("Kya")

index page

:Catalog Server

readIndexHeadings()

:Catalog Servlet

From member or nonmember page

index page is cached by browser

index()

Customer

sd U1

496 Appendix B

sd U2

Customer

:PCarModel[]

loop

:PCarModel Home

create(id,m,p)

p=getPrice()

m=getModelNumber()

id=getId()

[for each CarModel in list]

:Catalog Server

Figure B.36: Sequence diagram for U2:View Results

Extend with U3

results page

:Catalog Servlet :CarModel

Subsystem Design 497

sd U3

c2

:CarModel DetailsHome

c2=create(ms,e,p,d,a,t)

t=getPoster()

a=getAdvert()

d=getDescription()

p=getPrice()

e=getEngineSize()

ms=getMakes()

c1=findByCarModelId(id)

readCarModelDetails(id)

:Catalog Server c1:Car ModelDetails

Figure B.37: Sequence Diagram for U3:View CarModel Details

Extend with U7

details page

details(id)

Customer

:Catalog Servlet

:PCarModel DetailsHome

498 Appendix B

sd U4

ref

Server

Servlet

query(cs,ms,es)

search page

browser

is cached by

search page

page

:Category Home

:CarModel DetailsHome

List

list=findByQuery(q2:CatalogQuery)

q2 is a business version of q1

Figure B.38: Sequence diagram for U4:Search

U2(list)

:Make Home

List

findEngineSizes()

List<String>

findMakeNames()

List<String>

readCarModels(q1:PCatalogQuery)

int[]

readEngineSizes()

String[]

readMakeNames()

String[]

readCategoryNames() findCategoryNames()

:Catalog

:Catalog

From member or nonmember

search()

Customer

:CarModel Home

Subsystem Design 499

sd U5 :Authentication Servlet

member page

setSessionId(792)

getPassword()

a=isGoodMember()

Figure B.39: Sequence diagram for U5:Log On

U6, U8, U9, U10 or U12

Extend with

792

in browser session

792 stored

:Member Home :Internet Account

setSessionId(792)

getPassword()

m:Member

m=findByMembershipNumber("M1")

:Authentication Server

logon("M1","xyz",true)

logon("M1","xyz",true)

Member

500 Appendix B

sd U6

m2=create(n,a2,s,d,c2)

c1=getCard()

d=getAmountDue()

s=inGoodStanding()

a1=getAddress()

n=getName()

m1:Member

Figure B.40: Sequence diagram for U6:View Member Details

membership page

m2

a2 and c2 are protocol versions of a1 and c1

id retrieved from browser session

:Member Home

m1=findBySessionId(id)

:Membership Server

readMember(id)

:Membership Servlet

membership()

Member

:PMember Home

Subsystem Design 501

sd U7

details page

ok()

confirmation page

n

:Member Home :CarModel Home

n=getNumber()

create(m,c)

c=findByPrimaryKey(cid)

m=findBySessionId(id)

createReservation(id,cid)

session

id retrieved from browser

:Reservations Server

Figure B.41: Sequence diagram for U7:Make Reservation

confirmReserve()

confirmReserve page

reserve(cid)

Member

:Reservations Servlet :Reservation Home

:Reservation

502 Appendix B

sd U8

Figure B.42: Sequence diagram for U8:View Rentals

create(d,c2)

c1=getCar()

d=getDueDate()

:Rental Home

c2 is a protocol version of c1

List

list=findByMember(m)

[for each Rental in list]

PRental[]

loop

session

id retrieved from browser

:Member Home

m=findBySessionId(id)

:Rentals Server

readRentals(id)

:Rentals Servlet

rentals page

rentals()

Member

:Rental

:PRental Home

Subsystem Design 503

sd U9 :Membership Servlet

:Member Home

setPassword(n)

getPassword()

m=findBySessionId(id)

changePassword(id,o,n)

id retrieved from browser session

:Membership Server :Internet Account

setPassword(n)

m:Member

Figure B.43: Sequence diagram for U9:Change Password

password page

confirmChange()

confirmChange page

changePassword(o,n)

page

password

password()

Member

504 Appendix B

Extend with U11

sd U10

page

List

list=findUnconcludedByMember(m)

create(n,s,t,c2)

c1=getCarModel()

t=getTimestamp()

s=getState()

n=getNumber()

c2 is a protocol version of c1

[for each Reservation in list]

PReservation[]

loop

id retrieved from browser session

:Reservation Home

m=findBySessionId(id)

:Member Home :Reservation

Figure B.44: Sequence diagram for U10: View Reservations

reservations

:Reservations Server

readReservations(id)

:Reservations Servlet

reservations()

Member

:PReservation Home

Subsystem Design 505

sd U11

cancel()

getMember()

r=findByPrimaryKey(rid)

findBySessionId(id)

cancelReservation(id,rid)

:Member Home :Reservation Home

Figure B.45: Sequence diagram for U11:Cancel Reservation

reservations page

confirmCancel()

confirm cancel page

:Reservations Server

id retrieved from browser session

:Reservations Servlet

cancel(rid)

Member

r: Reservation

506 Appendix B

sd U12

home page

setSessionId(0)

:Internet Account

setSessionId(0)

m:Member

Figure B.46: Sequence diagram for U12:Log Off

id retrieved from browser session

Home

:Member

m=findBySessionId(id)

Server

Servlet

logoff(id)

:Authentication

:Authentication

logoff()

Member

Subsystem Design 507

508

Appendix B

B.6 CLASS SPECIFICATION This section documents the results of the speciﬁcation phase of the iCoot development. Since there are many classes in iCoot, speciﬁcations are only included for one ServerLayer class and one BusinessLayer class, as examples. Each speciﬁcation is given as comments in the source code.

B.6.1 Server Class Speciﬁcation ReservationsServer Class Specification Invariants: NONE // No invariants, because this is a stateless object Methods: /* * Create a reservation for the member with session identifier i * and the car model with identifier c * * Preconditions: * i != 0 * * * *

For mem = MemberHome.getInstance().findBySessionId(i), mem != null mem.isingoodStanding() mem.getAmountDue() == 0

* c != 0 * CarModelHome.getInstance().findByPrimaryKey(i) != null * * Postconditions: * A new Reservation has been created for Member with session * identifier i and CarModel with identifier c * * Exceptions: * * * */

PServerException (checked) thrown if the server has a problem IllegalArgumentException (unchecked) thrown if parameters are invalid

public void createReservation(int i, int c) throws PServerException; /* * Read all reservations for the member with session identifier i * * Preconditions: * i != 0

Class Speciﬁcation

*

MemberHome.getInstance().findBySessionId(i) != null

* * Postconditions: * result != null * result contains all unconcluded reservations for Member * with session identifier i * result is a new array, exclusive to the client * * Exceptions: * * * */

PServerException (checked) thrown if the server has a problem IllegalArgumentException (unchecked) thrown if parameters are invalid

public PReservation[] readReservations(int i) throws PServerException; /* * Cancel the reservation with identifier r for the member * with session identifier i * * Preconditions: * * * *

i != 0 For mem = MemberHome.getInstance().findBySessionId(i) mem != null r != 0

* * * *

For res = ReservationHome.getInstance().findByPrimaryKey(r) res != null res.getMember() == mem

* Postconditions: * For res = ReservationHome.getInstance().findByPrimaryKey(r) * res.isConcluded() * res.getReason().equals("Canceled by customer"); * * Exceptions: * PServerException (checked) thrown if the server has a problem * IllegalArgumentException (unchecked) thrown if parameters are * invalid */ public void cancelReservation(int i, int r) throws PServerException;

B.6.2 Business Logic Class Speciﬁcation Member Class Specification

509

510

Appendix B

Invariants: /* * Values named in invariants, preconditions and postconditions * are attributes, with a getter and an optional setter. Each * invariant is an extra precondition for the corresponding setter * and an extra postcondition for the corresponding getter. */ id is fixed after creation id != 0 number != null number.size() != 0 internetAccount != null address != null Methods: /* * Fetch the receiver’s id * * Preconditions: NONE * Postconditions: * result == id * Exceptions: NONE */ public int getId(); /* * Fetch the receiver’s number * * Preconditions: NONE * Postconditions: * result == number * Exceptions: NONE */ public String getNumber(); /* * Set the receiver’s number to n * * Preconditions: NONE * Postconditions: * number == n * Exceptions: NONE */ public void setNumber(String n);

Class Speciﬁcation

/* * Fetch the receiver’s internetAccount * * Preconditions: NONE * Postconditions: * result == internetAccount * Exceptions: NONE */ public InternetAccount getInternetAccount(); /* * Set the receiver’s internetAccount to ia * * Preconditions: NONE * Postconditions: * internetAccount == ia * Exceptions: NONE */ public void setInternetAccount(InternetAccount ia); /* * Fetch the receiver’s address * * Preconditions: NONE * Postconditions: * result == address * Exceptions: NONE */ public String getAddress(); /* * Set the receiver’s address to a * * Preconditions: NONE * Postconditions: * address == a * Exceptions: NONE */ public void setAddress(Address a); /* * Fetch the session identifier of the receiver’s internetAccount * * Preconditions: NONE * Postconditions:

511

512

Appendix B

*

result == internetAccount.getSessionId()

* Exceptions: NONE */ public int getSessionId(); /* * Set the session identifier of the receiver’s * internetAccount to i * * Preconditions: NONE * Postconditions: * internetAddress.getSessionId() == i * Exceptions: NONE */ public void setSessionId(int i);

B.7 OUTLINE TEST PLAN B.7.1 Introduction The testing of iCoot will be continuous, with the involvement of developers, peers, customers, the build team and the testing team. • Developers will test their artifacts as they produce them. • Customers will be involved in the veriﬁcation of high-level artifacts, acceptance testing and beta testing. • Peers will review the artifacts produced by developers. • The build team will be responsible for build testing after the ﬁrst increment. • The testing team will be responsible for coordinating the testing process, including the production and maintenance of this plan, the testing phase itself and system testing in particular. This test plan gives an overview of the testing that will be carried out, followed by details of the testing tasks for each phase of development. It doesn’t address the implementation of a prototype for iCoot. The development of a prototype will be under the supervision of project managers, conducted using rapid, informal methods.

B.7.2 The Impact of Spirals and Increments Within each spiral of development, testing will be carried out by the developers. Each artifact will be subject to peer review, with the proviso that formal peer reviews won’t be necessary

Outline Test Plan

for the ﬁrst spiral. For every spiral after the second, formal peer reviews will concentrate on changes that have been made to the artifacts, to avoid duplicated effort. Similarly, customer reviews should concentrate on changes that have been made since the most recent review. After each complete set of spirals, the testing team will take over to manage the testing phase before the release of the next increment. After the ﬁrst increment, regression testing will ensure that iCoot is at least as functional as it was after the previous increment.

B.7.3 Testing of Non-Code Artifacts Use cases and UML diagrams will be produced by the development team with input from customers. In the early spirals, members of the development team will review each other’s work and changes will be made immediately. During the last spiral before release of an increment, formal peer reviews will be held with colleagues who have relevant expertise but who are not directly involved with the project. These reviews will be used to certify the artifacts from use cases all the way through to class-based speciﬁcations. Developers, peers and project managers will be responsible for ensuring that artifacts remain consistent over time.

B.7.4 Code Reviews After the ﬁnal coding phase within each increment, formal code reviews will be held with peers who are not involved in the project itself. During these reviews, manual white-box testing and metric tools will be used to identify any refactoring that is needed.

B.7.5 Test-Driven Development During implementation of the design, programmers will perform continuous testing of their work with the help of JUnit. These programmer-developed tests will comprise unit tests at the class level and integration tests for all classes that the developer owns. Each developer will ﬁx faults in their own code before making the code public. At the end of each spiral, the developers will work together to perform integration testing and subsystem testing at a fairly informal level, again using JUnit, the aim being to eliminate as many faults as possible before the formal testing phase. The development team will ﬁx any faults that they discover at this point, before the next spiral or the testing phase, as appropriate.

513

514

Appendix B

B.7.6 Assertions As detailed later in this test plan, programmers will be expected to add assertions to their code which will be enabled during development. During the testing phase, assertions will be initially enabled to make it easier to identify faults. Once all the tests are successful, they will be run again with assertions disabled, to check that none of the assertions have side effects. It will also allow performance with and without assertions to be compared. For release, assertions will be disabled but retained in the code so that they can be reactivated to help diagnose failures. So that the live system is not compromised by the disabling of assertions, further steps will be taken to protect iCoot. In order to protect the servers from accidental or malicious attack when assertions are disabled, an application ﬁrewall will be implemented within the server layer. This ﬁrewall will largely enforce the server layer’s preconditions in such a way that the checks can’t be disabled. In order to reduce the amount of incorrect information that reaches the server layer, another application ﬁrewall will sit beneath the user interfaces (in the control layer) to reject invalid requests from the user. This client ﬁrewall will be based largely on the preconditions of the control layer, but again implemented in such a way that the checks can’t be disabled. In addition, the Web server will be placed in a de-militarized zone, sandwiched between two conventional Internet ﬁrewalls to frustrate typical Internet attacks. There will be two styles of client ﬁrewall. For HTML-based clients, the ﬁrewall will be implemented within the servlets and standard Web techniques for preventing invalid requests will be used (omitting invalid selections from Web pages and checking input data using JavaScript, for example). For GUI-based clients, the ﬁrewall will be implemented in the control layer and invalid requests will be prevented by using standard GUI techniques (such as disabling buttons that shouldn’t be used and replacing text entry with drop-down lists and spin buttons).

B.7.7 Testing Phase Once the spirals for each increment have been completed, the code will be handed over to the testing team. Before formal testing begins, the testing team will re-run the unit and integration tests that have been produced by the development team, to verify that there are no known faults. The testing team will use the class and subsystem speciﬁcations to help them produce subsystem test cases, with accompanying test procedures. The system test cases will be based on the system use cases. Each test case will comprise a number of individual tests, each with test name, test description, test procedure and expected results. The testing team will address the following requirements: • Load testing, at average and maximum loads. • Soak testing, to verify that there is no corruption or exhaustion of resources over time.

Outline Test Plan

• Stress testing, to conﬁrm that iCoot fails elegantly. • Security testing. Test automation will be used, wherever possible, to reduce the cost of testing. In parallel, the testing team will organize acceptance testing, with the help of in-house volunteers and carefully selected customer personnel. Installation testing will then be performed using a testbed that comprises a signiﬁcant subset of the target platforms. Finally before release, beta testing will be performed at selected customer sites. All known faults will be ﬁxed by the development team prior to release, wherever feasible. Performance metrics (for example, average transaction time) and style metrics (for example, method size) will be gathered. Any unacceptable performance must be ﬁxed prior to release. Style issues will be recorded for input to the next increment.

B.7.8 Documentation Testing After the ﬁrst spiral, the documentation team will start to produce manuals and training materials. These will be subject to peer review. After several spirals, during the explicit testing phase that takes place before the release of each increment, the documentation will again be tested. This testing will comprise peer review, acceptance testing and beta testing.

B.7.9 Build Testing Programmers will use a source code management tool to ensure that all code is kept in a central location. They will be required to check out any code that they intend to work on, in order to avoid duplicated or overlapping effort. The project managers will be responsible for deciding who works on what. The build team will be responsible for managing the code repository. In order to cope with developer absences, the code management tool will be conﬁgured to allow checked-out artifacts to be checked back in, under controlled conditions. After the ﬁrst increment has been released, all further development will be subject to nightly build tests, run by the build team. This will involve building the entire system and running a signiﬁcant subset of the system tests.

B.7.10 Test Documentation and Logging All test cases will be documented and kept in a test repository under the control of the testing team. As tests are carried out, test results will be recorded and added to the repository. In order to encourage developers to test their own code before the testing phase, they won’t have to record test failures in the test repository. Since the developers’ test cases will be written in Java (using JUnit), they will be stored in the code repository; therefore, there will be no need to add such test cases to the test repository.

515

516

Appendix B

During the testing phase, test failures will be added to the repository. The testing team will collaborate with project managers to ensure that ﬁxes are allocated to members of the development team and completed. It will be the responsibility of project managers to ensure that adequate JUnit test cases are developed and run frequently. For reporting purposes, the development team must sign off the integration testing that takes place at the end of each spiral, by adding a corresponding entry to the test repository.

B.7.11 Testing Activities by Phase Set out below are the testing activities that will take place within each phase of development. The requirement for formal peer reviews will be relaxed for the ﬁrst spiral. Peers will be selected on the basis of expertise. • Requirements phase The business use cases, user interface sketches, the use case diagram, system use cases, use case priorities, supplementary requirements and activity diagrams (where used) will be reviewed by developers, peers and customers. • Analysis phase The analysis class diagram, state machine diagrams (where used) and communication diagrams will be reviewed by developers, peers and customers. • System design phase The deployment diagram, technology choices, layer diagram, layer interaction policy, concurrency policy and security policy will be reviewed by developers and peers. • Subsystem design phase The design class diagrams; database schema; user interface design (for HTML access to servlets and JSPs, applets, applications, and interfaces for mobile devices); and sequence diagrams will be reviewed by developers and peers. • Speciﬁcation phase For each class, preconditions, postconditions and class invariants will be speciﬁed. These assertions will be reviewed by the developers and peers. Subsystem speciﬁcations may be used where appropriate. These will be tested in a similar way to class speciﬁcations. • Implementation phase The testing during this phase consists of three parts: – Adding assertions (as per the class speciﬁcation) to methods. Other types of assertion, such as checking loop termination and avoidance of impossible conditions, are optional, but recommended. – Creating JUnit test cases and JUnit test suites. There will be at least one test case (which is likely to involve some integration testing) per class and one test suite per package. – Performing code reviews, by developers and peers. • Testing phase The testing phase will be the responsibility of the testing team. The development of test cases and test procedures by the testing team will be reviewed by peers. Testing will comprise: – The JUnit tests, to ensure that all known faults have been ﬁxed.

Glossary

– Subsystem testing, based on any subsystem interfaces speciﬁed during the design and speciﬁcation phases. – System testing, based on use cases (functional testing and load testing at average load and maximum load). – Stress testing, to conﬁrm elegant failure (as deﬁned during the design and speciﬁcation phases). – Security testing (aggressive attempts to break into iCoot without authorization). – Acceptance testing, based largely on productivity metrics. – Metrics, based on system performance and coding style. – Documentation testing, with the help of end users and system administrators. – Installation testing, using a signiﬁcant subset of target environments. – Beta testing, at selected customer sites. • Maintenance phase The testing team will be responsible for managing the reporting and ﬁxing of faults discovered after release, with the help of project managers and the development team. Between increments, ﬁxes may be implemented, regression-tested and released at the discretion of the testing team and project managers. Regression testing will comprise: – The JUnit tests. – Subsystem testing. – System testing. – Installation testing. Feedback from customers about possible improvements to iCoot will be passed on to the project managers, with a view to incorporation in the next increment.

B.8 GLOSSARY Term

Deﬁnition

(Business object, system object, analysis object, design object) AddressHome (Design object)

Where a Member lives.

Address

Assistant

Auk

(Business actor, system actor)

(Business actor)

AukInterface

(Analysis object)

Home for creating and ﬁnding Address objects. An employee at a store who helps Customers to rent Car objects and reserve CarModels. The pre-existing system that handles Customer details, Reservations, Rentals and the Catalog of available CarModels. Boundary for accessing Auk.

517

518

Appendix B

AuthenticationServer

(Design object)

(Design object) AuthenticationServlet (Design object) AuthenticationServerHome

BusinessLayer

(Design layer)

Car (Business object, system object, analysis

object, design object) CarDetails (Analysis object, design object) (Design object) CarHome (Design object) CarModel (Business object, system object, analysis object, design object) CarModelDetails (Analysis object, design object) CarModelDetailsHome (Design object) CarDetailsHome

CarModelHome

(Analysis object, design

Controls the logging on and logging off of Members to iCoot. Home for creating an AuthenticationServer. Makes the AuthenticationServer accessible in an HTML page in a Web browser. Contains objects that convert the PersistenceLayer into clean object-oriented application objects. Instance of a CarModel for rent kept by a Store. Extra details of a Car, such as number plate and VIN. Home for creating and ﬁnding CarDetails. Home for creating and ﬁnding Cars. A model in our Catalog, available for reservation. Extra details about a CarModel, such as advert and poster. Home for creating and ﬁnding CarModelDetails. Home for ﬁnding and creating CarModels.

object) Catalog

(Business object)

CatalogQuery

(Design object)

(Design object) CatalogServer (Design object) CatalogQueryHome

(Design object) CatalogServlet (Design object) CatalogServerHome

Category

(Analysis object, design object)

A document describing CarModels available for rent. A Member’s speciﬁcation of CarModels that they’re interested in when searching the iCoot on-line Catalog; includes categories, makes or engine sizes. Home for creating CatalogQuery objects. Controls access to CarModels that can be browsed or reserved (Members only) over iCoot. Home for creating a CatalogServer. Makes the CatalogServer accessible in an HTML page in a Web browser. Classiﬁcation of a Car that helps Customers ﬁnd what they’re looking for, e.g. ‘‘Sports’’ or ‘‘Luxury’’.

Glossary

CategoryHome

Collectable

(Design object)

(Design object)

com::nowhere::business

com::nowhere::control

com::nowhere::micro

(Design package)

(Design package)

(Design package) (Design package)

com::nowhere::persistence

com::nowhere::protocol

(Design package)

(Design package) com::nowhere::server (Design package) com::nowhere::servlets (Design package) com::nowhere::rmi

(Design package) Concluded (Design object) com::nowhere::swing

CootBusinessServer

CootGUIClient

(Design node)

(Design node)

CootHTMLClient

(Design node)

cootschema.ddl

(Deployment artifact)

CootServer

(Design node)

(Business object, system object, analysis object, design object)

CreditCard

CreditCardCompany

(Business actor)

Home for creating and ﬁnding Category objects. A ReservationState indicating that a Customer has been informed about a matched Reservation but has not yet collected it. Package containing the BusinessLayer classes. Package containing the ControlLayer classes. Package containing the MicroLayer classes. Package containing the PersistenceLayer classes. Package containing the protocol classes, used by the ServletsLayer, RMILayer and ControlLayer for communication with the ServerLayer. Package containing the RMILayer classes. Package containing the ServerLayer classes. Package containing the ServletsLayer classes. Package containing the SwingLayer classes. A ReservationState indicating that a Reservation is ﬁnished because it was collected, canceled or timed out. A process hosting iCoot services on a CootServer. A Customer’s machine hosting a J2SE or J2ME GUI for accessing iCoot over RMI. A Customer’s machine hosting a Web browser to access iCoot. Script used to generate database tables for Coot. A machine hosting a WebServer and CootBusinessServer. Used for conﬁrming the credit-worthiness of Members; must not have expired for a Member to be in good standing. Company that issues CreditCards and conﬁrms validity.

519

520

Appendix B

CreditCardHome

(Design object) Customer (Business actor, business object, system actor, system object, analysis object, design object)

Home for creating and ﬁnding CreditCards. A person who pays money in return for one of our standard services.

DBMS

A process hosting a relational database management system. A machine hosting a DBMS. The department that deals with unpaid fees. A ReservationState indicating that a Reservation that was Collectable has timed out or been canceled; means that a Car is in the Reserved area that must be moved back to the display area. An Enterprise JavaBean; an object within a standard Java framework that can handle transactions, network access and database access, behind an Internet ﬁrewall; these come in two varieties, session beans (for remote access to business services) and entity beans (for automatic mapping of data to and from a database). The client-side code for accessing the ServletsLayer; provided by the standard HTML Web browser. The standard network layer that sits between the HTMLLayer and the ServletsLayer. Java enterprise archive containing the servlets, JSPs and EJBs used by CootHTMLClients and, eventually, CootGUIClients. Folder containing static HTML for the iCoot site. Standard Java Exception that indicates an attempt was made to send a message with invalid parameters. Details required for a Member to log on to iCoot plus a record of their logged-on status. Home for creating and ﬁnding InternetAccounts.

DBServer

(Business object) (Design object)

DebtDepartment Displayable

EJB

HTMLLayer

(Design layer)

HTTPCGILayer

icoot.ear

iCoot

(Design layer)

(Deployment artifact)

(Deployment artifact)

IllegalArgumentException

InternetAccount

(Design object)

InternetAccountHome

(Design object)

Glossary

J2EE J2ME J2SE JDBC

JDBCLayer

(Design layer)

JRMP JSP

Keys

(Business object)

LegalDepartment

License

(Business object)

LogonController

Make

(Business Actor)

(Analysis object)

(Analysis object, design object)

MakeHome

(Design object) (Analysis object,

MemberHome

design

Enterprise Edition of the Java 2 platform. Micro Edition of the Java 2 platform. Standard Edition of the Java 2 platform. A standard Java library that provides access to all relational databases in a uniform way. Layer that accesses a relational database from Java (within the EJB framework). The communication protocol used by RMI. Java Server Page; a dynamic web page containing Java code, to be executed by the server, alongside static HTML. For operating a Car; Customers have copies when they’re renting; Store keeps copies for available Cars and reserve copies for all Cars; Store has serial number for reproduction by Make if all copies are lost or broken. The department that deals with accidents in which a rented Car has been involved. A document that must be presented in order to rent a Car or as proof of identity when a NonMember makes a Reservation. iCoot controller that controls logging on and off by Members. Each Car has one or more Makes that manufactures it. Home for creating and ﬁnding Makes. Home for ﬁnding and creating Members.

object) MembershipCard

(Business object)

MembershipServer

(Design object)

(Design object) MembershipServlet (Design object) MembershipServerHome

MemberUI

(Analysis object)

A laminated document issued by a store to a Member as proof of membership. Controls access to a member’s details, such as Address and CreditCard, over iCoot. Home for creating a MembershipServer. Makes the MembershipServer accessible within an HTML page in a Web browser. iCoot boundary used by Members to access the system.

521

522

Appendix B

MicroLayer

NeedingRenewal

(Design object)

NonMember (Business actor, business object, system actor, analysis object, design object)

NonMemberUI

Notiﬁable

(Analysis object)

(Design object)

(Design object) (Design object) PCar (Design object) PCarHome (Design object) PCarModel (Design object) PCarModelDetails (Design object) PCarModelDetailsHome (Design object) PCarModelHome (Design object) PCatalogQuery (Design object) PCatalogQueryHome (Design object) PCategory (Design object) PCategoryHome (Design object) PCreditCard (Design object) PCreditCardHome (Design object) PMake (Design object) PMakeHome (Design object) PMember (Design object) PMemberHome (Design object) PRental (Design object) PRentalHome (Design object) PReservation (Design object) PReservationHome (Design object) PAddress

PAddressHome

Objects using the Java 2 Micro Edition to access the RMILayer from a pervasive device, such as a mobile phone or set-top box; reserved for future versions of iCoot. A ReservationState indicating that the Reservation has not been matched for a week and must be renewed if it is not to expire. A Customer whose identity and creditworthiness have not been checked and who, therefore, must provide a deposit to make a Reservation or surrender a copy of their License to rent a Car. iCoot boundary used by NonMembers to access the system. A ReservationState indicating that a Reservation has been matched to an available Car but the Member has not yet been notiﬁed. Protocol version of Address. Home for creating InternetAccounts. Protocol version of Car. Home for creating PCar objects. Protocol version of CarModel. Protocol version of CarModelDetails. Home for creating PCarModelDetails. Home for creating PCarModels. Protocol version of CatalogQuery. Home for creating PCatalogQuery objects. Protocol version of Category. Home for creating a PCategory. Protocol version of CreditCard. Home for creating PCreditCards. Protocol version of Make. Home for creating PMakes. Protocol version of Member. Home for creating PMembers. Protocol version of Rental. Home for creating PRentals. Protocol version of Reservation. Home for creating PReservations.

Glossary

PServerException

(Design object)

(Design object) Rental (Business object, system object, analysis object, design object) PServerExceptionHome

(Analysis object, design object) (Design object) RentalsServer (Design object) RentalHome

RentalServerHome

RentalsServlet

(Design object)

Reservation (Business object, system object, analysis object, design object) ReservationHome (Analysis object) ReservationServerHome (Design object) ReservationsServer (Design object) ReservationsServlet

ReservationsSlip

ReservationState

(Design object)

(Business object)

(Analysis object, design

object) ReservationStateHome

(Design object)

RMI

RMILayer

(Design layer)

ServerLayer

ServletsLayer

(Design layer)

An indication that one of the objects in the ServerLayer has been unable to complete a request because of, for example, a database problem. Home for creating PServerExceptions. A contract between Nowhere Cars and a Customer to keep one or more Cars for an agreed period; subject to late fees if Car is not returned on time. Home for ﬁnding and creating Rentals. Home for creating a RentalServer. Controls access to a Member’s Rentals over iCoot. Makes the RentalServer accessible within an HTML page in a Web browser. The reserving of a CarModel by a Customer. Home for ﬁnding and creating Reservations. Home for creating a ReservationServer. Controls access to Reservations for Members over iCoot. Makes the ReservationServer accessible within an HTML page in a Web browser. A slip detailing the membership number, car model, timestamp and number for a Reservation. The state of a reservation, e.g. Waiting or Concluded. Home for creating a ReservationState. Standard Java mechanism for sending messages to an object over a network. Converts the ServerLayer into simple objects that can be accessed by the SwingLayer or MicroLayer; reserved for future versions of iCoot. Contains objects to access iCoot over a network. The server-side objects that provide access to iCoot from an HTML Web browser.

523

524

Appendix B

Store

(Business actor, design object)

SwingLayer

(Design layer)

(Analysis object, design object) VendorHome (Design object) Vendor

VIN

Waiting

(Design object)

WebBrowser

WebServer

(Design node)

(Design node)

A Nowhere Cars site from which Cars can be rented, CarModels reserved and Catalogs browsed or obtained. Objects using the standard Java library for accessing the RMILayer from a Java GUI; reserved for future versions of iCoot. Company that supplies one or more Cars. Home for creating and ﬁnding Vendor objects. Vehicle Identiﬁcation Number; unique number issued by the licensing authority and appearing on a plate riveted to the Car’s body. A ReservationState indicating that the Reservation has been made over the Internet but has not yet been satisﬁed, canceled or expired. A process providing HTML access to a CootHTMLClient. A process providing server-side access to a CootBusinessServer from a Web browser.

Appendix C Summary of UML Notation Used This appendix summarizes the UML notation used in this book, organized by diagram type: • • • • • • • • • • • •

object diagram communication diagram (business level) activity diagram use case diagram class diagram (analysis level) communication diagram (analysis level) state machine diagram deployment diagram (network topology) package diagram deployment diagram (with processes, artifacts and manifestations) class diagram (design level) sequence diagram

The order corresponds to the order in which the diagrams were discussed in this book. It’s worth looking at them in order because, in order to avoid repetition and clutter, some notation is highlighted in earlier diagrams but not in later ones. Of course, UML comments can appear in any diagram – in many cases, it’s better to use comments than to resort to a complicated piece of UML notation, especially if that means looking at the speciﬁcation. Although UML makes no distinction between the two ﬂavors of class diagram, at design level and analysis level, a distinction is made here because the ﬁrst of each pair is simpler in style and therefore uses less of the available notation. In order to avoid confusion with respect to the Java code fragments, the primitive types used for diagrams in this book are Java primitives, such as int and boolean, rather than UML primitives, such as Integer and Boolean. As a matter of style, arrays are generally avoided in favor of collection classes (such as List). Some types of UML diagram were not needed during iCoot development and they are not covered here: component diagram, composite structure diagram, interaction overview diagram, and timing diagram. This appendix cannot comprehensively cover even the diagrams it does address. See [OMG 03a] and [Fowler 03] if you want to ﬁnd out more.

Summary of UML Notation Used

o:Class2

ro le

of o

instance specification

link

aClass4

object composition aggregation

navigability

:Class3 attribute1 = "Hello" attribute2:Class5 attribute3 = 39

Figure C.1: Object diagram

Actor

sequence number

st

fir

1: sa

es

m ge

Entity1 direction link

2:second message

Boundary Actor

Boundary

Entity2

Figure C.2: Communication diagram (business level)

527

528

Appendix C

initial pseudoaction edge

(Class) Action 1

action

fork

(Actor) Action 2

(Organization) Action 3 join

decision

[condition] guard (Department) Action 4

merge

final pseudoaction

Figure C.3: Activity diagram

Summary of UML Notation Used

Actor

association

SystemName U2 Abstract Use Case

U1 Use Case

keyword

U4 Optional Use Case

<<extend>> relationship

U3 Specialized Use Case

<>

U5 Included Use Case

<< t ex en d> >

comment

{condition} constraint

U6 Conditional Use Case

system boundary

Abstract Actor

Specialized Actor

Figure C.4: Use case diagram

529

530

Appendix C

AbstractClass1 attr1[0..1] attr2 attr3:Type

ClassA

operation1():Type operation2(param1,:Type,param2:Type) abstractOperation()

Comment

Stand-alone comment

Abstract Class2 inheritance (specialization)

Concrete ClassB

Concrete ClassC

<> ClassF

1

0..1

operation() <>

direction

association name multiplicity

2..10 ClassD

role of D

role of E

3

*

ClassE attribute

Association Class

Figure C.5: Class diagram (analysis level)

Summary of UML Notation Used

1. 1a :m es sa ge 2( )

:Entity

re su lt

=

assignment

guard

1: [condition] message1(10)

Actor

thread

b e) yp :T m ra pa e, yp :T 3( ge sa es :m

1. 2

nested sequence number

1 1.

iteration

*: m es sa ge 4( pa ra m 1)

:Boundary

result

control:Controller

Figure C.6: Communication diagram (analysis level)

531

532

Appendix C

internal trigger initial pseudostate

er gg tri 3

trigger 1

Initial State

er gg tr i

transition

State

trigger 4

2 Final State

final pseudostate

Figure C.7: State machine diagram

*

multiplicity

<<device>> MachineA

0..1

node

<<device>> AClient

3..15

<<device>> MachineB

communication path

2

<<device>> MachineC

Figure C.8: Deployment diagram (network topology)

Summary of UML Notation Used

Package1 Class1 Class2

Package2 Class3 Class4 Package3 dependency

namespace operator

Package4::Package5

Package6

Figure C.9: Package diagram

533

534

Appendix C

* <<device>> Client

<<ExecutionEnvironment>> ProcessA

Protocol

2

<<device>> Server

<<ExecutionEnvironment>> ProcessC

<<ExecutionEnvironment>> ProcessB

<<use>> ArtifactA

subnode

usage

<<artifact>> artifactb.xyz

<<manifest>> manifestation

AModelElement

Figure C.10: Deployment diagram (with processes, artifacts and manifestations)

Summary of UML Notation Used

package frame

pkg apackage

<> AnInterface Abstract ClassA

+abstractMethod() realization required interface/ provided interface

Concrete ClassB

-supplier 5..9

Supplier 1..*

AnInterface

2 collection field (non-standard notation)

-field

0..1

*

-parts:List

Part

ClassC

1

-field

Associate -privateField:Type -privateClassField:Type -Associate() <> ~packageMethod(param:Type) #protectedMethod():Type +publicMethod() +classMethod()

creation method (constructor)

Figure C.11: Class diagram (design level)

535

536

Appendix C

sequence diagram frame

sd diagram1

:ClassA

Actor

object lifeline

message1(param) creation

message1(param:Type)

ClassB("abc") result

:Type

message2(:Type) info=message2() loop frame

loop

[condition] message1()

activation bar

selfcall()

result:Type message2()

ref

diagram2(info)

stop

reference frame

Figure C.12: Sequence diagram

aB :ClassB

Bibliography

Alexander, C., Ishikawa, S., and Silverstein, M. (1977) A Pattern Language: Towns, Buildings, Construction, Oxford University Press, ISBN: 0-195-01919-9 Alur, D., Crupi, J., and Malks, D. (2003) Core J2EE Patterns: Best Practices and Design Strategies, Prentice Hall, ISBN: 0-131-42246-4 Ambler, S. (2003) Agile Database Techniques, John Wiley & Sons, ISBN: 0-471-20283-5 Beck, K. (1999) Extreme Programming Explained: Embrace Change, Addison-Wesley, ISBN: 0-201-61641-6 Beck, K. (2002) Test Driven Development, Addison-Wesley, ISBN: 0-321-14653-0 Bloch, J. (2001) Effective Java Programming Language Guide, Addison-Wesley, ISBN: 0-20131005-8 Bodoff, S., Green, D., Haase, K., Jendrock, E., Pawlan, M., and Stearns, B. (2002) The J2EE Tutorial, Addison-Wesley, ISBN: 0-201-79168-4, available at http://java.sun.com Bolton, F. (2001) Pure Corba, SAMS, ISBN: 0-672-31812-1 Booch, G. (1993) Object-Oriented Analysis and Design with Applications, Benjamin Cummings, ISBN: 0-805-35340-2 Bustard, D., Kawalek, P., and Norris, M. (2000) Systems Modeling for Business Process Improvement, Artech House, ISBN: 1-580-53050-8 Campione, M., Walrath, K., and Huml, A. (1998) The Java Tutorial Continued: The Rest of the JDK, 2nd Edition, Addison-Wesley, ISBN: 0-201-48558-3, available at http://java.sun.com Campione, M., Walrath, K., and Huml, A. (2000) The Java Tutorial: A Short Course on the Basics, 3rd Edition, Addison-Wesley, ISBN: 0-201-70393-9, available at htttp://java. sun.com

Clark, A., and Warmer, J. B. (2002) Object Modeling With the OCL: The Rationale Behind the Object Constraint Language, Springer-Verlag, ISBN: 3-540-43169-1 Cockburn, A. (2000) Writing Effective Use Cases, Addison-Wesley, ISBN: 0-201-70225-8 Cockburn, A. (2001) Agile Software Development: Software Through People, Addison-Wesley, ISBN: 0-201-69969-9 Constantine, L., and Lockwood, L. (1999) Software For Use: A Practical Guide to the Models and Methods of Usage-Centered Design, Addison-Wesley, ISBN: 0-201-92478-1 Fowler, M. (1996) Analysis Patterns: Reusable Object Models, Addison-Wesley, ISBN: 0-20189542-0 Fowler, M. (2003) UML Distilled: A Brief Guide to the Uniﬁed Modeling Language, 3rd Edition, Addison-Wesley, ISBN: 0-321-19368-7

Bibliography

Gamma, E., Helm, R., Johnson, R., and Vlissides, J. (1995) Design Patterns: Elements of Reusable Object-Oriented Software, Addison-Wesley, ISBN: 0-201-63361-2 Grand, M. (1999) Patterns in Java, Volume 2, John Wiley & Sons, ISBN: 0-471-25841-5 Grand, M. (2002) Patterns in Java: Catalogue of Reusable Design Patterns Illustrated with UML, Volume 1, 2nd Edition, John Wiley & Sons, ISBN: 0-471-22729-3 Jacobson, I., Booch, G., and Rumbaugh, J. (1999) The Uniﬁed Software Development Process, Addison-Wesley, ISBN: 0-201-57169-2 ¨ Jacobson, I., Christerson, M., Jonsson, P., and Overgaard, G. (1992) Object-Oriented Software Engineering: A Use Case Driven Approach, Addison-Wesley, ISBN: 0-201-54435-0 Joy, W., Steele, G., Gosling, J., Bracha, G. (2000) Java Language Speciﬁcation, 2nd Edition, Addison-Wesley, ISBN: 0-201-31008-2, available at http://java.sun.com Kay, A. (1972) ‘A Personal Computer for Children of All Ages’, Proceedings ACM National Conference, August , Boston Larman, C. (2001) Applying UML and Patterns, Prentice Hall, ISBN: 0-130-92569-1 Lea, D. (1999) Concurrent Programming in Java: Design Principles and Patterns, 2nd Edition, Addison-Wesley, ISBN: 0-201-31009-0 Martin, J., and Odell, J. (1998) Object-oriented Methods: A Foundation – UML Edition, Prentice Hall, ISBN: 0-139-05597-5 McConnell, S. (1998) Software Project Survival Guide, Microsoft Press, ISBN: 1-57231-621-7 Meyer, B. (1990) Eiffel: The Language, Prentice Hall, ISBN: 0-13-247925-7 Meyer, B. (1997) Object-Oriented Software Construction, 2nd Edition, Prentice Hall, ISBN: 0-13-629155-4 Myers, G., Sandler, C., Thomas, T., and Badgett, T. (2004) The Art of Software Testing, 2nd Edition, John Wiley & Sons, ISBN: 0-471-46912-2 Object Management Group (2003a) UML 2.0 Superstructure Speciﬁcation, ptc/03-08-02, available at www.omg.org Object Management Group (2003b) UML 2.0 OCL Speciﬁcation, ptc/03-10-14, available at www.omg.org

Object Management Group (2004) Common Object Request Broker Architecture: Core Speciﬁcation, Version 3.0.3, OMG, formal/04-03-12, available at www.omg.org Raggett, D., Lam, J., Alexander, I., and Kmiec, M. (1997) Raggett on HTML 4, 2nd Edition, Addison-Wesley, ISBN: 0-201-17805-2 Robinson, M., and Finkelstein, E. (2004) Jakarta Struts for Dummies, John Wiley & Sons, ISBN: 0-764-55957-5 Rumbaugh, J., Blaha, M., Premerlani, W., Eddy, F., and Lorensen, W. (1991) Object-Oriented Modelling and Design, Prentice Hall, ISBN: 0-13-630054-5 Singh, S. (2000) The Code Book: The Science of Secrecy from Ancient Egypt to Quantum Cryptography, Anchor Books/Doubleday, ISBN: 0-385-49532-3

539

540

Bibliography

Taylor, D. (1997) Object Technology: A Manager’s Guide, 2nd Edition, Addison-Wesley, ISBN: 0-201-30994-7 Weaver, P., Lambrou, N., and Walkley, M. (2002) Practical SSADM 4: A Complete Tutorial Guide, Prentice Hall, ISBN: 0-273-65575-2 Wirfs-Brock, R., and McKean, A. (2002) Object Design: Roles, Responsibilities and Collaborations, Addison-Wesley, ISBN: 0-201-37943-0 World Wide Web Consortium (1999) HTML 4.01 Speciﬁcation, REC-html40, available at www.w3c.org

World Wide Web Consortium (2003) SOAP Version 1.2 Part 1: Messaging Framework, REC-soap12-part1-20030624, available at www.w3c.org Yergeau, F., Bray, T., Paoli, J., Sperberg-McQueen, C. M., and Maler, E. (1999) Extensible Markup Language (XML) 1.0, 3rd Edition, W3C, REC-xml-20040204, available at www.w3c.org

Index

abnormal path 143, 155, 427 abnormality 106 abstract classes 59–62 Abstract Factory design pattern 346 abstract method 59 abstract use case 150 abstraction 58, 229 abuses 80–81, 225, 375 acceptance testing 418–19 access 222 accessibility 90 accessor message 277 accessors 277–8 accidental boundary blurring 206 accommodation 212 across-method testing 441–2 action 143 activation bar 312 ActiveX 223, 242 activity diagram 118, 143–5, 528 see also iCoot case study actor 135, 149–50, 191, 450–51, 460 see also business actor, system actor Ada 4, 206 Adapter design pattern 354–5 adapting patterns 367–70 adding implementations to class hierarchy 58–9 adding operations to classes 194 address 214 advantages of copies 308 advantages of inheritance 68 advantages of multi-threading 318–19 advantages of proxies 308 advantages of servlets 253 aggregation 21–3, 171–7, 279 see also association agile methodology 5, 116 alarm bells 184 Alexander, Christopher 328

alias 330 alien radio signals 217 alpha testing 413 amber light 160–62, 207 ambiguity 374 analysis 102–3, 466–76 analysis paralysis 107 analysis-level class diagram 121–2, 530 analyzing the problem 166–203 dynamic analysis 188–96 overview of the analysis problem 169–70 static analysis 170–88 why do analysis? 168–9 Ant 415 anthropomorphism 14 anti-patterns 328 Apache project 415 Apple 213 applet 230–2, 242–4, 246, 301, 353 plus RMI 230–2 application builder 109 application ﬁrewall 394–5, 398–9, 402, 427–8 code 398 application server 257 ARPAnet 214 array 33, 41, 179, 207, 274 artifact testing 117 artifacts 263, 408, 423 artiﬁcial constraint 220 ASCII 214 ASP 247 assembler 4 assembly language 4 assertion checking tools 422 assertion mechanism 402 assertion syntax 384–5 assignment 89 assignment of priorities 162

542

Index

association 21–3, 171–8, 291–4 mapping 291–4 see also aggregation association classes 182–3, 283–4, 293–4 association labels 177–8 atomic value 293 attribute 16–18, 169, 178–82, 278–9, 467 identical or equal? 16–18 mapping 278–9 or relationship? 180–82 see also iCoot case study; relationship attribute event 233 attribute list 178 Auk 142, 159, 226, 451 authentication 222, 224, 248–9 authorization 116, 222 automated theorem proving 278 automatic traceability 425 automating testing 421–2 autonomous subsystem 225 back-end conﬁgurations 258 batch 208 bean 256 Beck, Kent 430 behavior 14, 42, 158, 375 behavior state machine 196 behavioral pattern 330 behavioral testing 410 beneﬁts of three-tier architecture 211–12 Berners-Lee, Tim 214 best practice 108, 396 beta testing 413–14, 426 binary chop search 358, 429 binary numbers 4 birth of a system 133–5 bit strength 224 black-box testing 169, 409–410, 412, 426 black-box system testing 414 see also use case testing blurred boundaries 206, 207 Booch, Grady 115 boolean 41, 179 bottleneck 243 boundary 142, 189, 191–2 boundary blurring 206, 207

boundary conditions 427 brainstorming 108, 115, 158–9 breadth of inheritance hierarchies 421 brevity 277 brief history of patterns 328–30 software patterns today 329–30 brief history of programming 4–5 broadband 214 broadcast 233 browser 214, 242 bubble 148 bugs 12, 105, 408–9 build team 416 build testing 415–17, 515 see also iCoot case study build tools 415 business actor 137–8 see also actor business logic class speciﬁcation 509–12 business logic tier 211 business model 136–45 identifying business actors 137–8 identifying business use cases 139–41 illustrating use cases on an activity diagram 143–5 illustrating use cases on a communication diagram 141–3 writing project glossary 138–9 business modeling 102 business process modeling 136 business requirements 132–3, 450–56 business requirements modeling 135, 168 business rules 221 business service 228, 272, 306–314, 483 see also iCoot case study business service realization 312–14 see also iCoot case study business use cases 139–41 C# 4, 206 C++ 4, 34, 41 application and browser language 243 existence of composition 176 history 206 inheritance 68, 73 see also garbage collection

Index

cache 358 Caesarian Cipher 223 calculation faults 428 callback 337 can of worms 179 cancer research 218 candidate class 169 candidate key 290 casting 89 catch an exception 397 cathode ray tube 209 certiﬁcate authority 224 certiﬁcation 425 CGI see common gateway interface CGI-with-scripts 246, 250–2 CGI-with-servlets 252–3 challenge mechanism 311 change history 116 chaos 100, 112, 208 checked exception 255, 400 checking conditions dynamically 396–7 choosing networked system topology 208–20 client–server versus distributed architectures 216–18 depicting network topology in UML 218–20 extranets and virtual private networks 215–6 history of network architectures 208–10 Internet and World Wide Web 214 intranets 215 network computers 213–14 personal computers 212–13 three-tier architecture 210–12 choosing technologies 240–69 back-end conﬁgurations 258 client tier technologies 242–4 client tier to middle tier protocols 244–5 Java e-commerce conﬁguration 258–62 middle tier technologies 246–7 middle tier to data tier technologies 247–8 other technologies 248–50 typical front-end conﬁgurations 250–58 UML packages 262–6 see also iCoot case study clarity 368 class comment 396

class constant 401 class diagram 34, 118, 121–2, 168–9, 466, 530, 535 drawing 172–3 see also iCoot case study class hierarchy design 55–8 class interface speciﬁcation 372–405 class invariant 383 class naming style 139 class parameter 91 class relationships 171 class scope 382 class speciﬁcation 103, 508–12 class survey 176 classes 34–7, 278–9 mapping 278–9 what they deﬁne 37–9 classical phases in software production analysis 102–3 deployment 104 design 103 implementation 104 key questions 105–6 maintenance 104–5 requirements 102 speciﬁcation 103 testing 104 classiﬁcation 35 classifying business services 309–311 client code 237, 307 client tier 211 client tier technologies 242–4 client tier to middle tier protocols 244–5 client–server architecture 216–18 client–supplier interaction 30–32 closed layer 227 COBOL 4, 206, 243 code accessing 225 code book 224 code coverage 420 code cracking 223 code reuse 43–7 code review 424, 513 code sharing 61 collaboration 27, 30–32, 195, 293, 353

543

544

Index

collections 55 combining methodologies 113–14 combining patterns 367–70 comment 176–8, 396 commission faults 409 commit transaction 315 common design patterns 331–66 see also design patterns common gateway interface 245, 250 disadvantages 252 Common Object Request Broker Architecture see CORBA communication diagram 27, 118, 122–3, 141–3, 168, 527, 531 elements of 192–4 see also iCoot case study compatibility 88 compile-time concept 265, 400 compiler 4 compiler switch 401–2 completing the Store class 442–5 complex life cycle 367 complexity 70, 104, 150, 170, 195 component diagram 118 component testing 414–15 Composite design pattern 361–4 composite structure diagram 118 composition 66–8, 73, 171–7, 278–9 mapping 278–9 versus inheritance 68–9 compound key 290 computer modeling 417 concrete classes 61–2, 68 concrete method 59 concurrency 208, 220–2, 315–16, 482–3 see also iCoot case study concurrency control 272 conﬁgurable copies 309 conﬁgurable proxies 309 conﬂict of interest 316 connections 21, 116 consistency checking 117 constrained genericity 92 constraint 153, 222

constructor 37 context 88–9, 356 continuous testing 406–445 automating tests 421–2 example of test-driven development using JUnit 431–45 preparing for tests 422–3 test-driven development 430–31 testing strategies 424–6 testing terminology 408–411 types of test 411–21 what to test for 426–9 contracts 389–91 documenting 396 enforcement 394–5 violation 397–8 control of access 222 control of polymorphism 83 see also inheritance controller 189, 191–2, 230 controlling multiple tasks 317–18 controlling multiple threads 318–19 conversion 89 cookies 260 copies 307–9 advantages of 308 disadvantages of 308–9 CORBA 115, 255–6, 257, 258 CORBA naming services 246 correctness 117 corruption 319 coupling degree 421 CPPUnit 430 crashing 376 creating test suite 439–41 creation expression 16 creational pattern 330 creative nature of software development 111 criteria for testing 426–9 CRT see cathode ray tube cryptography 223–5 customers 108 data deﬁnition language 287 data dictionary 138 data management 210

Index

data manipulation language 287 data query language 287 data structure theory 429 data tier 211 data tier technologies 247–8 data type 279 database access 315 database management system 211, 226–9, 287–8 DBMS see database management system DDL see data deﬁnition language deadlock 237, 319 debugging 294, 350, 374 decision 144 declarative language 287 declarative programming 381 decryption 215, 223–5 defect 409 deﬁciencies 105 deﬁnition of an object 13–16 deﬁnition of speciﬁcation 375–6 degree of cohesion 421 degree of coupling 421 delegation 66–9 dependency 85, 154 depicting network topology in UML 218–20 depicting objects 19–20 deployment 104 deployment diagram 118, 123–4, 218–19, 481–3, 532, 534 see also iCoot case study deployment survey 219 depth of inheritance hierarchies 421 derived attribute 179 design 103 Design by Contract 103, 385–95 application ﬁrewalls 395 contracts and inheritance 389–91 enforcing the contract 394–5 reducing error-checking code 391–4 see also Eiffel; Meyer, Bertrand design patterns 46, 331–66 Adapter design pattern 354–5 Composite design pattern 361–4 Facade design pattern 352–3

Factory Method and Abstract Factory design patterns 346 Flyweight design patterns 358–61 Iterator design pattern 342–6 Multiton design pattern 341–2 Observer design pattern 331–7 Proxy design pattern 364–6 Singleton design pattern 337–41 State design pattern 347–52 Strategy and Template Method design patterns 356–8 design priorities 207 design-level class diagram 124–5, 535 designing business service 306–314 business service realization 312–14 classifying business services 309–311 session identiﬁers 311 using proxies and copies 307–9 designing a class hierarchy 55–8 designing for concurrency 220–22 designing for security 222–5 digital encryption and decryption 223–5 general security rules 225 designing the solution 204–445 choosing technologies 240–69 continuous testing 406–445 designing the subsystems 270–325 designing the system architecture 204–39 reusable design patterns 326–71 specifying the interfaces of classes 372–405 designing subsystems 270–325 designing business services 306–314 ﬁnalizing user interfaces 300–306 handling multiple activities 315–20 handling persistence with relational database 286–300 mapping analysis class model into design class model 275–86 transactions 314–17 using patterns, frameworks and libraries 314 designing system architecture 204–39 choosing a networked system topology 208–20 design priorities 207

545

546

Index

designing system architecture (Continued) designing for concurrency 220–22 designing for security 222–5 partitioning software 225–37 steps in system design 207–8 desktop publishing 243 details of system use case 155–6 developer model 145–62 identifying system actors 146–7 identifying system use cases 147–9 preconditions, postconditions and inheritance 156–8 prioritizing system use cases 159–62 specializing actors 149–50 supplementary requirements 158 system use case details 155–6 use case relationships 150–55 use interface sketches 158–9 development tools 116–17 developmental phase testing 424–5 digital certiﬁcate 224 digital decryption 223–5 digital encryption 222, 223–5 digital signature 225 disabling dynamic checks 400–402 implementing a compiler switch 401–2 using assertion mechanism 402 disadvantages of CGI 252 disadvantages of copies 308–9 disadvantages of inheritance 68 disadvantages of proxies 308 discovering patterns 367–70 disgruntled customer 221 distributed architecture 216–8 distribution of messages 57 DML see data manipulation language documentation testing 419, 517 see also iCoot case study documenting a contract 396 domain model 136 donkey work 104 downcast 91 DQL see data query language drawing class diagram 172–3 drawing object diagrams 172–3

drawing relationships 173–8 association labels, roles and comments 177–8 multiplicity 175–7 drawing use case realizations 189–90 dumb terminal 209, 210 duplication of effort 328 dynamic analysis 169, 188–96 adding operations to classes 194 boundaries, controllers and entities 191–2 communication diagram elements 192–4 drawing use case realizations 189–90 responsibilities 194–5 state modeling 195–6 dynamic analysis model 168 dynamic artifact 121, 123, 126 dynamic binding 85–7 dynamic casting 91 dynamic checking 380–82, 389, 394 dynamic condition checking 396–7 dynamic modeling 115 dynamic type systems 80–81 e-commerce 301, 426 e-mail 242, 243, 245, 246, 317, 365, 426 eating an elephant 108–9, 112, 113 eavesdropping 225 effectiveness 117 efﬁciency 211 Eiffel 4, 34, 40–41, 384–5 application and browser language 243 creation of 383 history 206 informal speciﬁcations 380 inheritance 68–9, 73 see also Design by Contract; Meyer, Bertrand EJB 46, 247, 256–8, 314 elegant failure 386, 418 elements of communication diagram 192–4 elephant eating 108–9, 112, 113 embedded systems 426 enabling dynamic checks 400–402 implementing a compiler switch 401–2 using assertion mechanism 402 encapsulation 20–21, 43, 275, 289, 309 encryption 223–5 end-of-method checks 397

Index

end-to-end development tool 116 enforcing contracts 394–5 Enterprise Java Beans see EJB entity 142, 191–2, 290–91 entity bean 257 entry point 32 enumeration 341 environment testing 418 see also installation testing equality 16–18, 427 error recovery 386 error-checking code 381, 386, 388, 394 ethical hackers 225, 420 event source 233 events 233–7, 249 exception catching 397 exception handling 380 exception throwing 397 exceptions 388 explicit cases 89–91 expression evaluation 88 expressiveness 191 extensibility 162 eXtensible Markup Language 249 extension 152 external clients 398–9 external suppliers 399–400 external systems 398–400 external clients 398–9 external suppliers 399–400 extra capacity 206 extranet 215–6 extreme programming see XP extrinsic state 359 Facade design pattern 352–3 factory 314, 367 Factory Method design pattern 346 fail-safe code 378, 387 failure 408–9 faking identiﬁers 311 fat client 214 fault ﬁxing 12, 106 faults 12, 105, 107, 408–9 faults of commission 409

faults of omission 409 feedback 191 ﬁeld visibility 274–6 ﬁle server 209 ﬁle transfer 244, 245 ﬁnal state 195 ﬁnalizing user interfaces 300–306 avoid multiple windows 304–6 be guided by use cases 301 keep it simple 301–2 use notebooks 302–3 use wizards 303–4 ﬁnding classes 170–71 ﬁrewall 215 ﬁx 409 ﬁxed values 319 ﬂexibility 212, 223, 311 ﬂight simulator 218 ﬂow chart 143 ﬂy-by-wire aircraft 376–7, 417 ﬂying as polymorphic message 84 Flyweight design pattern 358–61 foreign key 291–4 forks and joins 144 formal speciﬁcation 376–8 in OCL 383–4 Fortran 4, 206, 243 frame 313 frameworks 314, 422 front-end conﬁgurations 250–58 CORBA 255–6 EJB 256–8 HTML/CGI-with-scripts 250–52 HTML/CGI-with-servlets 252–3 RMI 254–5 FTP daemon 246 function libraries 45 functional programming 4 functional speciﬁcation 102 functional testing 410 Gamma, Erich 430 Gang of Four 328 garbage collection 32–4, 66, 176, 427 gathering requirements 130–65 birth of a system 133–5

547

548

Index

gathering requirements (Continued) business perspective 136–45 developer perspective 145–62 use cases 135–6 general security rules 225 generalization 36, 62, 69 generic message 91 genericity 91–3, 207 see also templates getter 37, 275, 331, 335, 350 getting it right 186–7 getting it wrong 183–5 glass-box testing 411 see also white-box testing global error variable 388 global intranet 215 global scope 382 global sign-on 249 glossary see project glossary good objects 188 graceful failure 386, 418 graphical user interface 32, 68 graphs 23–5 see also trees green light 160–62, 207 green screen 209 gremlins 420 grid computing 218 guard 144 guess work 376 GUI see graphical user interface guideline on polymorphism 87–8 guidelines for inheritance use 73 gurus 46, 106 hackers 215, 223, 225, 311, 420 handler code 388 handling multiple activities 315–21 controlling multiple tasks 317–18 controlling multiple threads 318–19 thread safety 318–21 handling persistence with relational database 286–300 database management systems 287–8 mapping associations 291–4 mapping entity classes 290–91

mapping object state 294–300 relational model 288–90 hidden attribute 20 hierarchy 21–3 high cohesion 47 high-level language 4 high-throughput programming logic 211 hijacking 223 history of network architectures 208–10 home 191, 284, 314 host machine 210 how object-oriented program works 32 HTML see HyperText Markup Language HTML/CGI-with-scripts 246, 250–52 HTML/CGI-with-servlets 252–3 HTTP see HyperText Transfer Protocol hybrids 309, 367, 384 hyperlink 214 HyperText Markup Language 214, 229 HyperText Transfer Protocol 214 IBM 117, 212–13, 377 see also Vienna Development Method icon 191 iCoot case study 450–526 analysis 466–76 business requirements 450–56 class speciﬁcation 508–12 glossary 517–24 outline test plan 512–17 subsystem design 483–508 system design 476–83 system requirements 456–66 identifying business actors 137–8 identifying business use cases 139–41 identifying class relationships 171 identifying system actors 146–7 identifying system use cases 147–9 identity 16–18, 427 identity attribute 187 illustrating use cases 141–5 on activity diagram 143–5 on communication diagram 141–3 immutability 318–19 imperative code 381 imperative programming 106

Index

implementation elements 58 implementation inheritance 68–9 implementation priority 159–62 implementations 58–9, 104 implementing the Car class 434–5 implementing compiler switch 401–2 implementing a stack class 64–9 implementing a stack using composition 66–8 implementing a stack using inheritance 65–6 inheritance versus composition 68–9 implementing a stack using composition 66–8 implementing a stack using inheritance 65–6 implicit casting 89 inclusion 152 incremental methodology 112–13 impact of increments 512–13 see also iCoot case study independent nodes 218 indexed ﬁle 287 industrial espionage 215 inﬁnite recursion 428 inﬂexibility 111 informal speciﬁcation 376, 378–80 in Eiffel 384–5 in Java 396–402 information retrieval 233 Information Super-Highway 214 inheritance 36, 53–77, 156–8, 297–300, 389–91 abstract classes 59–62 adding implementations to a class hierarchy 58–9 and contracts 389–91 control of polymorphism 83 designing a class hierarchy 55–8 guidelines for use of 73 implementing a stack class 64–9 multiple inheritance 69–73 redeﬁning methods 63–4 and specialization 156–8 inheritance versus composition 68–9 initial state 195 initialization 236, 348, 428

installation testing 418 instant messaging 242–3, 245 instruction 233 intangible objects 183–7 getting it right 186–7 getting it wrong 183–5 integer attribute 291 integration testing 412–13 integrity 222, 224 intentional boundary blurring 206 inter-process communication 385–6 interaction overview diagram 118 interface 42, 238, 300–306, 372–405 internal implementation 56 Internet 117, 212, 214, 301 searching 316 see also intranet Internet ﬁrewall 215 Internet Service Provider 214 intranet 117, 215, 243 see also Internet intrinsic state 359 introduction 2–8 about this book 5–8 background 4 brief history of programming 4–5 methodologies 5 intuition 313 investment protection 212 invoking an operation 28–30 irrefutability 222, 224 ISP see Internet Service Provider iteration 192 iterative methodology 109, 111–12 Iterator design pattern 342–6 Jacobson, Ivar 115, 135, 148, 188, 191 jargon 138 Java 2 Enterprise Edition 46, 314 Java 2 Micro Edition 314 Java 2 Standard Edition 254, 314 Java 4, 6, 34 application and browser language 243 explicit casting 90 inheritance 69, 73 static type system 81

549

550

Index

Java DataBase Connectivity 231, 247, 287 Java e-commerce conﬁguration 258–62 Java informal speciﬁcation 396–402 checking conditions dynamically 396–7 documenting a contract using comments 396 enabling and disabling dynamic checks 400–402 external systems 398–400 signaling contract violations using RunTime Exceptions 397–8 Java layers 230–32 Java Messaging Service 237, 249 Java Naming and Directory Interface 254 Java Remote Method Protocol 244 Java synchronization 319–21 JavaBeans guidelines 110 JDBC see Java DataBase Connectivity JMS see Java Messaging Service JNDI see Java Naming and Directory Interface JRMP see Java Remote Method Protocol JSP 249 JUnit 425, 430 JUnit and test-driven development 431–45 completing the Store class 442–5 creating a test suite for regression testing 439–41 implementing the Car class 434–5 refactoring tests 435–9 testing across methods 441–2 testing the Car class 433–4 Kay, Alan 12 ‘Keep the bar green to keep the code clean’ 443 key questions concerning software development 105–6 keys 223, 284, 290 keyword 286, 339, 383 KLOC 420 knowledge 233 language translation 367–8 latency reduction 101, 319

laws of physics 106 layer diagram 477–9 layer interaction policy 263, 479 layers 208, 227–30, 317 for single-tier systems 228 transactions in 317 translation layers 229–30 for two- and three-tier systems 228–9 lazy initialization 338, 341 legacy system 212 legal ﬁeld value 374 legal message 196 libraries 314 life-savers 112 lifeline 312, 350 light coupling 332 lightweight copies 309 link 25–7, 173 see also navigability link table 293 listener 233–7, 246 load balancing 211 load testing 417–18 soak testing 417 stress testing 418 loading tools 422 lock 320 locking 316 logging 515–16 logic programming 4 logical partitioning 212 see also partitioning software London Underground 25 loops 428 loose coupling 195, 421 machine code 4 machine-to-machine messaging 249 mainframe 208, 423 maintenance 104–5 manageable chunks 208 manipulation 191 manual checking 424 many-to-many association 282–3, 292–3 mapping associations 291–4 association classes 293–4

Index

many-to-many associations 292–3 one-to-many associations 292 one-to-one associations 291–2 mapping entity classes 290–91 mapping inheritance 297–300 mapping object model to relational model 290 mapping object state 294–300 mapping inheritance 297–300 mapping operations 273–4 mapping other types of association 279–84 margin of error 106 maximum possible error 381–2 memory address 285 memory error 427 memory leak 33, 427 message comment 396 message digest 225 message ﬂow 335 message ﬂow in layers 233–7 message ﬂow using events 237 message postcondition 383 message precondition 382–3 message-driven bean 257 messages 27–8, 251 polymorphic 83–5 metaclass 40, 339 method size 421 methodologies 5 metrics 420–21, 426 metrics-gathering tools 422 Meyer, Bertrand 384, 388 see also Design by Contract; Eiffel middle tier 211 middle tier technologies 246–7 middle tier to data tier technologies 247–8 midi computer 209, 210 mini computer 209, 210 mini–midi combination 209, 210 mission statement 5, 134, 450 misuse of values 80 mix-in inheritance 69–70, 72–3 mnemonics 4

mobile phone 212 mock object 412 Modula 4, 206 monitor 322 monoliths 210 moving from analysis to design 273–84 accessors 275–6 mapping classes, attributes and compositions 278–9 mapping operations 273–4 mapping other types of association 279–84 universal identiﬁer 284–6 variable types 275 visibility of ﬁelds 275–6 MT-hot code 318, 320 MT-safe code 318–22 multi-branch statement 427 multi-threading 318–20, 353 advantages for client 318–19 advantages for server 319 multi-valued attribute 293 multiple inheritance 69–73 multiple task control 317–18 multiple thread control 318–19 multiple windows 304–6 multiplicity 175–7 multitasking 317–23 Multiton design pattern 341–2 multiuser access control 116 mutability 319 see also immutability mutual exclusion 318–20

name clash 70–71 naming service 254 narrow context 89 NASA 219 natural language 141, 153, 374 navigability 25–7 see also links navigable association 282 need for development tools 116–17 negative testing 418 nested transactions 315 network architectures 208–10

551

552

Index

network computers 213–4 network programming 224, 365 network proxy 330 network trafﬁc 308–9 networked operation 117 networked system topology 210–22 new methodologies 110–114 combining methodologies 113–14 incremental methodology 112–13 iterative methodology 111–12 spiral methodology 110–111 new programming styles 109–110 nightly build 415 nodes 25, 220 non-code artifact testing 513 non-null parameter 384 non-programmers 329 nontrivial work 214 normal path 143 normal path code 388 notebooks 302–3 nuclear power station 376 null value 279, 281, 383 nullable column 292, 296 numerical methods 381 object concepts 10–51 association and aggregation 21–3 classes 34–7 deﬁnition 12–13 depicting objects 19–20 encapsulation 20–21 example collaboration 30–32 garbage collection 32–4 graphs and trees 23–5 how an object-oriented program works 32 identical or equal? 16–18 invoking an operation 28–30 links and navigability 25–7 messages 27–8 reusing code 43–7 shared data and shared operations 39–40

terminology 41–3 types 40–41 what does a class deﬁne? 37–9 what is an object? 13–16 Object Constraint Language see OCL object diagram 19, 118, 172–3, 527 Object Management Group 115 object orientation 63 object state 294–300 object to relational model mapping 290 object-oriented extension 287 object-oriented language 63, 382 object-oriented methodologies 114–17 need for developmental tools 116–17 UML, RUP and XP 115–16 object-oriented paradigm 12 object-oriented programming 4 Object-Oriented Programming, Systems, Languages and Applications conferences see OOPSLA conferences object-oriented speciﬁcation 382–5 formal speciﬁcation in OCL 383–4 informal speciﬁcation in Eiffel 384–5 objects and transactions 316–17 Observer design pattern 331–7 OCL 120, 153, 377, 383–4 formal speciﬁcation 383–4 omission faults 409 ‘one size ﬁts all’ computer 211 one-tier architecture 208–10 one-time pad 224 one-to-many association 281–2, 293 one-to-one association 280–81, 291–2 OOPSLA conferences 329 open layer 227 operation invocation 28–30 operation list 194, 467–9 operator 311, 427 opinion 367 optimistic concurrency 315–16 see also concurrency optimization 276 origins of UML 115–16

Index

other technologies 248–50 other types of association 279–84 association classes 283–4 many-to-many associations 282–3 one-to-many associations 281–2 one-to-one associations 280–81 overview of analysis 169–70 package diagram 118, 264, 533 package visibility 276, 432, 480–81 packages in UML 262–6 pair programming 116 panic 100 parameter passing 89 parameters 28 paranoia 61, 388, 395 parent class 54 partition 144 partitioning software 225–37 Java layers: applet plus RMI 230–32 layers 227–30 message ﬂow in layers 233–7 systems and subsystems 226 see also logical partitioning part–whole hierarchy 21–2, 36 Pascal 4, 34, 206 password 225, 427 pattern language 328–9 pattern template 330–31 pattern use 314, 367 patterns 314 PDA see personal digital assistant peer review 425 peer-to-peer architecture 217–18 performance 275 performance improvement 211 performance monitoring tools 422 pernicious faults 107 persistence 286–300 persistence layer 230, 317 persistent data 272, 287 personal computers 212–13 personal digital assistant 212, 244 pessimistic concurrency 315–16 see also concurrency

Petri net 143 pictorial programming language 115 piggybacking 245 PL/1 206 platform testing 418 see also installation testing pointers 26, 90 polymorphic messages 83–5 polymorphic variables 82–3 polymorphism 82–5, 229, 389 polymorphic messages 83–5 polymorphic variables 82–3 polymorphism guideline 87–8 pool 358 port 246 post-release testing 426 postcondition 377, 383 of specialization 155–8 precedence 427 precondition 337, 382–3 of specialization 155–8 preparing for testing 422–3 previous faults 429 primary key 290–91, 293–4, 297, 316 prime number 218, 224 primitives 40–41, 80, 90, 178, 276 priorities of design 207 prioritizing system use cases 159–62 priority 317 privacy 222, 224 private attribute 20 private inheritance 68–9 private visibility 63, 276 problem analysis 166–203 process-level concurrency issue 221 production prototyping 109 project glossary 138–9 proof-of-concept 109–110, 168 proprietary applications 242 protected visibility 274 protection 225 protocol 42 protocol state machine 196 prototyping 109, 168

553

554

Index

proxies 307–9, 364, 367 advantages of 308 disadvantages of 308 Proxy design pattern 364–6 public attribute 382–3 public visibility 63, 274 pure event 234 quality of software 116 quantum cryptography 224 Quantum theory 29, 376 RAD see rapid application development rapid application development 109–110 Rational Application Developer 117 Rational Corporation 115, 117 Rational Uniﬁed Process see RUP real numbers 377 real subject 364 real world data 423 realization 236 reasons for analysis 168–9 reasons for request failure 28–9 records 33 recursive state see recursivity recursivity 361 red light 160, 162, 207 redeﬁning methods 63–4 redeﬁnition 58, 85 reduced latency 101 reduced redundancy 116 reducing error-checking code 391–4 redundancy 237 refactoring 408–9, 436 refactoring tests 435–9 regression testing 419, 439–41 relational database and persistence 286–300 relational model 288–90 keys 290 mapping object model to relational model 290 tables 289 relations 289 relationship 169, 173–8, 180–82 or attribute? 180–82 drawing 173–8 see also attribute

remote login 242 remote method invocation see RMI repeatability 101 repeated inheritance 71–2 requirements capture 102, 125, 145 responsibility 194–5 result code 388 reusable design patterns 326–71 common design patterns 331–66 discovering, combining and adapting patterns 367–70 history of patterns 328–30 pattern template 330–31 using patterns 367 reuse repository 45 reusing code 31, 43–7, 68, 88 reverse engineering 288 right machine for job 211 Ripple 5, 117–26 class diagram (analysis level) 121 class diagram (design level) 124 communication diagram 122–3 deployment diagram 123–4 sequence diagram 125–6 summary 446–9 use case diagram 120–21 risk management 328 RMI 230–32, 244, 254–5 RMI registries 246 RMI-over-IIOP 256 robustness 428 rogue program 215 roles 137, 177–8 Rose, Rational 117 Rumbaugh, James 115 run-time system 34, 68, 71–2, 80–81, 319–21 RunTimeException 397–8 RUP 5, 101–2, 115–16, 192 Russian doll 361 safety 222 safety-critical system 376, 378, 417, 426 scenario 313, 414 scenario testing 414 scheduler 317 schema 287, 290–91, 374

Index

search for extraterrestrial intelligence 217–18 secure HTTP 244 secure sockets layer 244 security 208, 222–5, 482 improvement 212 testing for 419–20 self-taught developer 132 semantic error 384 semantic gap 287 sensitivity 225 separation of concerns 211 sequence diagram 118, 125–6, 142, 536 sequence numbers 27 server class speciﬁcation 508–9 server object 272, 309 server tier 211 servlet 32, 247 advantages 253 session bean 257 session identiﬁers 311 set-top box 212, 305 SETI@home 217–18 setter 39, 277, 331, 350, 352 setting the scene 9–128 inheritance 52–77 object concepts 10–51 software development methodologies 98–128 type systems 78–97 shared data 39–40 shared operations 39–40 shifting marketplace 113 signaling contract violation 397–8 signing off 425 Simple Object Access Protocol (SOAP) 249 simplicity 301–2 single inheritance 69 single log-on 311 single sign-on 249 single-tier system layers 228 Singleton design pattern 337–41 Smalltalk 4, 12, 20, 34, 40 dynamic type system 80–81 history 206, 329 inheritance 69, 72

patterns 328–9 test-driven development 430 smoke test 415–16 soak testing 417 SOAP see Simple Object Access Protocol software debugging 374 software development methodologies 98–128 classical phases in software production 102–6 new methodologies 110–114 object-oriented methodologies 114–17 Ripple overview 117–26 software engineering and the waterfall method 106–110 software engineering 106–110 see also waterfall methodology software patterns today 329–30 software reliability 28 Solaris 209 specialism 12 specialization 36, 69, 150, 158 specializing actors 149–50 speciﬁcation 103, 372–405 speciﬁcation testing 409 see also black-box testing; use case testing speciﬁcation-exercising tools 422 specifying class interface 372–405 Design by Contract 385–95 dynamic checking 380–82 formal speciﬁcation 376–8 informal speciﬁcation 378–80 informal speciﬁcation in Java 396–402 object-oriented speciﬁcation 382–5 what is a speciﬁcation? 375–6 spiral methodology 108, 110–111 impact of spirals 512–3 see also iCoot case study sponsor 110–112, 133, 137–40, 145–8, 158 SQL see structured query language square pegs and round holes 90 SSADM 5 SSL see secure sockets layer stack class 64–9 start transaction 315 state 314 State design pattern 347–52 state machine diagram 118, 195, 532

555

556

Index

state machine survey 195 state modeling 195–6 state pattern 294 state attribute 296 static analysis 169–88 association classes 182–3 attributes 178–82 drawing class and object diagrams 172–3 drawing relationships 173–8 ﬁnding classes 170–71 good objects 188 identifying class relationships 171 tangible versus intangible objects 183–7 static analysis model 168 static artifact 120–2 static modeling 115 static type systems 80–81 steps in system design 207–8 stop 350 strategies for testing 424–6 Strategy design pattern 356–8 strengthening postconditions 390 stress testing 418 strings 25 strong cohesion 195, 421 strong coupling 195, 421 structural pattern 330 structural testing 411 see also white-box testing structured methodologies 5 structured programming 4 structured query language 287 structures 33 stubs 412 style checker 422 stylish code 105 subclass 59–66, 72–3, 83 subsystem design in iCoot 483–507 business service realization 494–507 business services 483 BusinessLayer class diagram 487–8 BusinessLayer ﬁeld list 488–492 database schema 494

protocol objects class diagram 492–4 ServerLayer class diagram 485–6 ServerLayer ﬁeld list 486 ServerLayer message list 486–7 ServletsLayer class diagram 483–4 ServletsLayer ﬁeld list 484 ServletsLayer message list 484–5 user interface design 494 subsystem partition 208 subsystems 226, 270–325 Sun 244 SUnit 430 superclass 54–6, 59–64, 71–3, 86 supplementary requirements 158 Swing library 230–31, 302 synchronization 280, 283–4, 319–21 system actors 146–7 system architecture 204–39 system design 103, 476–83 system design steps 207–8 system malfunction 409 system requirements 132–3, 456–66 system requirements modeling 102, 135, 152, 168 system topology 207 system use case details 155–6 system use cases 147–9 details 155–6 prioritizing 159–62 systems 226 tables 289 tadpoles 27 tangible objects 183–7 getting it right 186–7 getting it wrong 183–5 TCP/IP 214, 244 technology choices 207, 240–69, 476–7 see also iCoot case study teletype unit 208, 209, 210 Telnet daemon 245, 246, 248 template 91–3, 330 headings 330–31 pattern template 330–31 see also genericity Template Method design pattern 356–8

Index

terminology 41–3, 408–411 test automation software 421–2 test bed 418, 422 test case 423 test data 423 test documentation 515–6 test harness 423 test plan 422 test preparation 422–3 test procedure 423 test suite 423, 431, 441 test types 411–21 test-driven development 116, 408, 430–31, 513 see also iCoot case study testing 104 testing across methods 441–2 testing after release 426 testing the Car class 433–4 testing during development 424–5 testing during testing phase 425–6 testing frameworks 422 testing of non-code artifacts 513 testing for security 419–20 testing strategies 424–6 after release 426 during development 424–5 during testing phase 425–6 see also preparing for testing testing terminology 408–411 black-box testing 409–410 white-box testing 410–411 thin client 214 thousands of lines of code (KLOC) 420 thread safety 272, 318–21 ﬁxed values 319 immutability 318–19 synchronization in Java 319–21 thread-safe code 320–21 threads 317 threads of execution see threads three amigos 115–16 three-tier architecture 208, 210–12 three-tier system layers 228–9 throw an exception 397

throw it over the wall syndrome 12–13 tight coupling 195, 421 time out 337 time-slicer 208, 317 time–space trade-off 59 timing diagram 118 tolerances 106 traceability 116 trade-off 59, 368 trafﬁc lights 160–62, 207 transaction time 420–21 transactions 314–17 guidelines for using transactions with objects 316–17 pessimistic and optimistic concurrency 315–16 in upper layers 317 transactions in upper layers 317 transformation 368 transition 195 translation layers 229, 231 transparency 309 trees 23–5, 429 see also graphs trigger 195 trivial program 331 trivial type 171 tunnel 215 tweaking 62 two-tier architecture 209–10 two-tier system layers 228–9 type casting 88–9 type systems 78–97 dynamic binding 85–7 dynamic and static type systems 80–81 explicit casts 89–91 genericity with templates 91–3 polymorphism 82–5 polymorphism guideline 87–8 type casting 88–9 types 40–41, 178 types of relationship 171 types of test 411–21 acceptance testing 418–19 alpha testing 413

557

558

Index

types of test (Continued) beta testing 413–14 build testing 415–17 component testing 414–15 documentation testing 419 installation testing 418 integration testing 412–13 load testing 417–18 metrics 420–21 regression testing 419 testing for security 419–20 unit testing 412 use case testing 414 typical front-end conﬁgurations 250–57

use case testing 412, 414 see also black-box testing use cases 135–6, 301 use of copies 307–9 use of proxies 307–9 USENET 242 user interface layer 228 user interface sketches 158–9, 300, 456–9 see also iCoot case study user interfaces 300–306 using assertion mechanism 402 using patterns 314, 367 using transactions with objects 316–17 utility message 191

UML depicting network topology 218–20 notation 527–37 origins of 115–16 packages 262–6 unchecked exception 255, 400 uncrackable encryption algorithm 224 understanding the problem 129–203 analyzing the problem 166–203 gathering requirements 130–65 Uniﬁed Modeling Language see UML Uniform Resource Indicator 214 unit testing 412 unit of work 314 universal identiﬁer 284–6, 297 Unix 208, 213, 227, 422 unreachable code 428 unusual situations 427 upcast 91 urban planning 328 URL see Uniform Resource Indicator use case diagram 118, 120–21, 147, 460–61, 529 use case guidance 301 use case illustration 141–5, 451–66 on activity diagram 143–5, 453 on communication diagram 141–3, 452 use case model 135 use case realizations 169, 189–90, 312, 470–76 see also iCoot case study use case relationships 150–55 use case survey 148–9, 461–2

validation 409 value 136 variable types 274 variables: polymorphic 82–3 VDM see Vienna Development Method vehicle identiﬁcation number 290 veriﬁcation 409 versioning 116 Vienna Development Method 377 VIN see vehicle identiﬁcation number violation 381 violation of contract 397–8 virtual poker game 222 virtual private network 215–16 viruses 215 visibility of ﬁelds 274–5 VPN see virtual private network waterfall methodology 106–110 see also software engineering weak cohesion 184, 421 weakening preconditions 390 Web server 214 WebStart 254 what a class deﬁnes 37–9 white-box testing 410–412, 426 wide context 89 widget 159 wind-tunnel test 417 window system 210 wizards 303–5

Index

work in progress 413 workﬂow analysis 136 workings of object-oriented program 32 workstation 209, 213 World Wide Web 214 writing project glossary 138–9

XML see eXtensible Markup Language XP 5, 115–16 xUnit 430 Z 377 zooming 178

559